Maybe the client signed with another agency, or your boss gave the raise to a coworker instead.

The truth is that people see one bad looking slide and assume your desk is buried in clutter, your car is full of wrappers, and your whole life’s a disaster.

But the 5 Claude Design prompts below create files that look so good you won’t have this problem. Instead, people will think “daaaaang, this one has their stuff together.”

With a simple copy / paste you’ll be able to build:

• Decks

• Docs

• Landing pages

• Wireframes

• Animations

The outputs will be tidy, jaw-droppingly gorgeous, and on brand. And as a bonus, I’ll even show you how to build an app icon (if you should need something like that).

You’ll get all of that, and more, in the prompt pack below.

None of us want to admit it, but we judge others based on how good their decks and documents look. A slide with clashing colors or a bad layout says something about a person’s taste, tidiness, and attention to detail. And when opportunities come around, they don’t automatically go to the best, sometimes they go to the one who looks the best.

That’s why the prompts I’m sharing with you today are so important.

Think of them like a digital dress suit for your files.

If you use these prompts, your clients will think you’re a pro, your boss will ask you to teach them how you did it, and you’ll be the envy of your colleagues. Or at least, that’s the hope.

You don’t have to be a designer or a coder to make this work for you. If you can slide a mouse around and move your fingers, you can produce amazing looking files, I promise.

To make it even easier, I recorded a video walking through EXACTLY how to use each prompt.

So, without further ado, enjoy 👇

Before we begin

Two things:

1. Every prompt below assumes you already built a design system and saved it in Claude Design. If you don’t know how to do that yet, check out this article (it’s totally free and only takes 15 minutes).

2. Attach it. Each of the prompts below start out by creating a new design in Claude Design and attaching your design system. This is KEY to making everything else work. If you want your designs to have the same fonts, colors, gradients, etc (whatever your brand looks like), this step is non-negotiable.

After that, copy any prompt below and fill in the [brackets] with whatever you want. Ramble, be messy. You don’t have to be precise, you just gotta be in the ballpark.

Don’t forget the basic skill here: you brief it with what you want, it builds, and then you pick from whatever variations it spits out. You’re more like a director, not a designer.

Claude Design Prompt for Decks

This one is for pitch decks, sales decks, investor decks, internal updates, presentations, etc.

Watch the video walkthrough:

Read more

Last week I built a brand almost as good in Claude Design for just $200 a month.

In this article, I’m going to show you how I did it.

I know it seems too good to be true. You’re probably thinking, “Sure, but isn’t AI design just slop?”

Fair.

We can all spot AI slop from a mile away. I hate it as much as you do. But with this new method, the designs are downright gorgeous and definitely NOT slop.

You’ll see that by just typing a few simple sentences you can create designs so good that people will think you paid thousands for it.

And don’t worry… I’ll walk you through everything step-by-step. That way you’ll be able to do this too even if you’re not a designer (like me).

Oh, and you don’t even need the pricey $200 Claude plan I’ve got. The $20 plan does everything I’m about to teach you.

All you really need to make my method work are opinions.

Surely you’ve got those, right?

Okay, good Let’s do this 👇

Subscribe now

You know it’s wrong, but have no idea why

I’ve got pretty good design taste, and I bet you do too.

(Even if you don’t realize it).

Have you ever looked at something and been like, “Wow, this is amazing!” but didn’t know why exactly? You could just feel it?

Or have you ever used an app and been like, “This design sucks bad. I don’t know why. But it does.”

Well, congrats. That’s taste.

And to the dismay of designers everywhere, notice what’s missing from both of those reactions. There’s no mention of:

• Composition

• Values

• Colors

• Spacing

• And other fancy design terms

In other words, we can tell you quite emphatically that something isn’t working, but we’ll never be able to tell you how to fix it.

For designers, this basically means we are clients from hell 🤣

(Go us! 🥳)

Now, I want to give designers their due. If you’ve ever spent any time in Canva trying to hack a graphic together, you know that design is a real skill, and it’s shockingly hard to do well.

There are about ten million tiny decisions hiding inside any design that make it look great...

And every single decision is a chance to totally screw the whole thing up.

This is why designers exist. The rest of us mortals eventually get frustrated and toss our laptops out the window. But if you’ve been down this path, you also know that designing something in Canva vs. designing something with a pro is usually just trading one frustration for another.

Here’s what I mean:

Once upon a time, I ran a weekly live stream for 5 years straight. Every week, we’d need a YouTube thumbnail for it, and that thumbnail also doubled as the title card for the stream itself.

So, it was important.

Like clockwork, my talented design team would put something together and send it over. I’d take one look and go, “No, that’s not quite right.”

“Okay, what do you want us to fix?” they’d ask.

And I’d just sit there for a few minutes trying to pinpoint why I didn’t like it.

Coming up empty handed, I’d eventually blurt out, “Man, I don’t freaking know what to tell you to fix. I don’t even know what’s wrong with it. All I know is it’s not good. It doesn’t feel right. Can you just make me like, 10 variations on the concept and I’ll pick the best one?”

But of course, they couldn’t do that because we didn’t have the time to churn that many variations out. So, I’d settle for a design I didn’t love and that was the loop I lived over and over.

Me: knowing it was wrong.
Them: waiting on direction I couldn’t give.

I was stuck, they were stuck, and the image still wasn’t right. It was maddening for all of us.

But now, with Claude Design, this is no longer a problem.

It’s like a world-class designer with the heart of a golden retriever

I’m sure you’ve heard of vibe coding by now.

It’s where any normal non-technical person can use AI to build software. The general idea is that you just describe what you want in plain English, and AI writes all the code.

You just kinda steer it with feedback as it goes.

Well, I’m coining something new:

Vibe Branding

(original, I know)

It’s basically the same thing as vibe coding, but with - you guessed it - branding.

You describe the vibe you want in plain English. Claude Design does the rest.

The coolest part about vibe branding is that you can just blab on a keyboard and Claude Design, with it’s infinite patience, will do it’s best to interpret whatever it is you want and make you happy.

It’s kinda like having a world class designer with the heart of a golden retriever.

If you want a brand that’s loud, well by golly, it’ll do loud.

If you want something that’s high energy, it’s got you covered.

If you want something that’s slightly ephemeral with a touch of nostalgia and a hint of goth in the fresh pouring rain... well... whatever the heck that is... it’ll do that too.

And every time you have a new request or revision, Claude Design will meet you at the door with a wagging tail and abounding joy.

All you need are messy opinions, and it just works.

Beyond that, with the method I’m going to teach you in a minute, everything it cranks out is visually gorgeous.

In my experience, if Claude Design hands me four options, it’s not a thing where there are three duds and one keeper that I have to rescue.

No. They’re all friggin’ awesome. Any of the four would work (examples coming later).

That’s because all those big brain designer skills, like understanding:

• Composition

• Contrast

• Spacing

• Color harmony

• Balance

• Et cetera

A lot of these components are objective, which means they can be defined and written down as rules. The AI models today are so good at those rules now they’ll make your head spin.

So, Claude Design just spits fire.

But there’s one thing Claude Design doesn’t have:

It doesn’t know WHAT FEELING you’re going for.

It can hand you four gorgeous versions and still not know which one is right for your thing, your mood, the way you want somebody to feel when they see it.

That part is still yours.

All you have to do is show up with an opinion.

That’s how easy this is now... like, do-it-from-the-couch-in-your-pajamas-easy.

So, let’s do a mini-recap:

• Claude Design is a design agency in a box, sitting right there on your laptop

• It works the way a good agency works, but comes with a 99.9% discount

• You riff, it comes back with a smorgasbord of designs

• The designs rock because Claude has taste built into it

• You just follow your opinions and mood

That’s the world we’re in now. You get a whole design agency - a good one - in a box for pennies.

I hope you’re psyched. I’m psyched.

Let’s keep going 👇

Subscribe now

All you have to do is type stuff

Now, I’ll show you how to do this too.

I’m going to walk you how I built a design system for an app I built called YadaYada. I think it’s beautiful:

I’ll explain everything start to finish, so you can turn around and do the same thing for your own project.

The best part is you’re not going to open a single design tool. No Canva, no Photoshop, none of that.

All you’re going to do is type.

Step 1: Pop over to Claude Design

Click this: claude.ai/design

There’s no download, and nothing to install. Claude Design is just a web app.

Congrats, you completed step one.

Awesome, I know.

Quick note before we go further, because this confuses some folks:
Claude Design is not an image generator. It’s not DALL·E or Nano Banana or Midjourney or the ChatGPT image thing. You’re not making pictures of puppies and rocket ships riding pink unicorns. Claude Design can build real artifacts. Think slide decks, wireframes, websites, etc. In this tutorial, we’re building what designers call a “design system”. It’s the actual colors and fonts and rules of a brand that make all of those other things look pretty.

Step 2: Prompt it in plain English

For all you Claude Design veterans out there, this is where I detour from conventional wisdom.

Most tutorials will tell you to hop into Claude’s design system builder. I don’t like to start out that way, and think that’s how you get a platter of AI slop.

So, stay away from this button for now:

Instead, I start riffing on designs straight from here and go to the design system builder later:

Here’s the only real skill I need you to have, and it barely counts as a skill.

(Sorry to all you English majors, I’m just being honest 😬)

You have to be able to describe what you want like you’re telling a friend over coffee. There’s no need to be precise. You just have to be in the ballpark.

So, ramble off the kind of brand you’re going for, the colors you lean toward, whether you want it warm or cold, high energy or calm. Try to describe the feeling you want to invoke. This is the level you want to riff at. You don’t need hex codes, but if you’ve got them, great. Same with design jargon or an actual spec... bring them if you have them, but nobody’s asking you to.

Here’s the exact prompt I used for YadaYada.

I’m actually a tinge embarrassed to share this, because most of you think I’m some kind of AI wizard and I’d like to keep it that way.

But no.

I’m just like anybody else. I ramble, I repeat myself, I’m not always the most clear in what I want. And this STILL WORKED.

(IE. You can do this, too!)

This is my exact prompt that I voice-dictated into Claude Design:

Hey, I need you to help me build a new design system. This one is going to be for an app called Yada Yada. This is a voice dictation app and the way it works is it uses local models on your machine and doesn’t go to the cloud. That way you don’t have to pay a subscription basically. Now, I like design systems that are kind of minimal in nature, easy to use, but also I like bright, fun, high energy. I like neon colors. I like pastel colors. I like things that are, you know, not really corporate. So I like fun, bright, high energy, things that feel fresh. I like lots of white space with splashes of color. And I definitely like pastels and I like neon colors and I’d like to kind of lean into those. So can you help me build? You know, I need a few different options for design system for the app. I don’t want you to design the app, right? I don’t want you to come up with how the app could look, you know, with the button placement and all that stuff right now. I just want to focus on the design system itself, like the fonts, the colors, the tokens, et cetera.

Notice what I did NOT do there.

I didn’t pick colors. I didn’t type a single hex code. I gave it a vibe and told it I wanted a few different options.

Step 3: Answer its questions

After this prompt, Claude Design will think for a while and then ask you a handful of questions.

This part is easy to get through because the questions are multiple choice. There’s almost never anything open-ended.

It’ll ask stuff like “how many directions do you want to explore?” with options like, one, two, three, or “you pick for me”.

You won’t get visuals yet, it’s still just gathering information. I answered a few questions on whim and let it decide the rest.

It takes like, 5 minutes tops.

You really can’t screw this part up.

Step 4: Iterate and pick a winner (the fun part)

Now it gets fun. After cranking for 15-20m, it’ll lay out a few complete directions and gives each one a name.

There was one called Sherbert I liked right away. Beautiful fonts, beautiful colors, most of the way to what I wanted.

So, I just told it that:

I love the gradient on Sherbert. Can I see maybe three more variations on Sherbert? I’d like to see some different gradient options. Also, I’m not positive that’s my favorite font on Sherbert, but it is my favorite font out of all the others you shared.

A few minutes later I had three siblings sitting next to Sherbert.

One called Aurora stood out to me, so again, I just told it:

Aurora is beautiful. Love the fonts. Love the gradient temperature. Let me see 3 more just to see if you can top Aurora

It cranked for a while, then spit out a variation that I just thought was stunning. It called this one Opal.

I can’t tell you how delightful this process actually is.

You get to see as many directions as you want laid side by side, and it costs you nothing but some tokens and the time it takes to look.

Step 5: Build your design system MVP

After finding a winner, I simply prompted Claude like this:

Ohmy. Opal is beautiful. Let’s do that one

Remember, back in Step 1 I told Claude the real goal was a brand kit. Everything since has just been nailing the look so it could go build it.

So, after you pick your winner, stand back and let Claude cook.

I told it I wanted Opal, and to keep it to the foundations for now... the colors, the type, the spacing, the way the gradient works, the brand itself... and to NOT run ahead and start designing the actual app.

I just wanted the layer everything else gets built on top of.

So that’s what it made.

A stack of token files with every color and font and size in them, the whole foundational kit, sitting in one place. Here’s what it looked like when it was finished:

One thing I noticed is that Claude Design likes to spread the system across a bunch of separate pages. This probably comes down to personal preference, but I wanted mine all on one tall page I could open and scan in a single scroll.

So I just prompted it to build one scrollable overview page with the whole system laid out top to bottom, and to save that as standalone HTML.

Step 6: Export that baby

Once it’s done and you’ve got the tweaks the way you want them, click “Share” in the top right, and then “Project HTML”.

After this, you’ll see a pop-up asking if you want to download the “Project archive” or “Standalone HTML”. Both of these options work, but I prefer a standalone HTML file so that’s what I went with

After this, it’ll cook a bit longer, and then give you a link to download straight to your machine. The HTML is great because it works offline, fonts and all baked right in. It’s also really easy to use as a reference file for any future work you do.

Step 7: Build your FULL design system

But wait... THERE’S MORE!

We’re still not done yet, but if you wanted to stop here you could. You just built out a pretty lightweight MVP of a design system. However, if you want the full experience, I’d encourage you to continue.

Now, we’re going to go back to Claude Design, but this time we’re going to click the “Design systems” tab, and then the “Create design system button.”

Next, you’ll get a popup with two options. You’re going to click “Create here”

This will present another screen with some questions on it. You can tell Claude a little bit about your company if you want.

The most important, thing, however, is to click the “Add fonts, logos and assets” button here and upload the HTML file you just downloaded:

Note: If you opted to download the “Project Archive” earlier, you can choose the “Link code from your computer” option and point it at the archive on your computer.

After that, click “Continue to generation” in the top right of the screen, and Claude Design will get to cooking on a FULL design system workup.

When it’s finished, it should look something like this. It’s basically a fleshed out version of the MVP:

Note: You probably saw an app icon fly by in the screen. This was not created by Claude Design and not covered in this tutorial.

Okay, time to pause for a second. Let’s think about what just happened.

You shot from the hip, you vibe branded (as the cool kids say), and you’ve got a complete, genuinely professional design system sitting in a file.

It’s not the same thing a $200k brand agency would produce, and I don’t want to pretend it is. An actual agency would build out an entire strategy, perform in-depth customer research, interview stakeholders and host workshops. But for nailing brand colors, tone, energy, typography, and everything else... it’s definitely like, 80% as good and covers the overwhelming majority of use cases.

And that took what? 15-20 min tops?

And you didn’t need to know the first thing about design to make it.

This design system is super high leverage. EVERYTHING else you create after this can now ride this design system, and the outputs will come out looking on brand.

Think of all the digital assets you produce:

• Decks

• Charts

• Docs

• Sites

• Videos

• Infographics

• etc.

All you have to do is attach your design system, and ask AI to help you create the asset so it looks like your brand.

It’ll take care of the rest.

Subscribe now

Welcome to 2026

Just a few years ago, this would have cost thousands of dollars minimum, not including the time you would have lost in the back and forth on the project.

You just did it with AI in a browser tab for whatever your subscription cost is.

And the cool part I want to make sure you catch is that it’s not like your skill level rose. You didn’t get better at design. On the contrary, the bar just dropped all the way down to something you’ve been able to do since you were a kid:

Have opinions.

Think about that for a minute.

If you have opinions you can build things now.

The world just opened wide.

Just writing these words I’m getting excited, a little itchy to go dive in and build something else.

If you found this useful, it’d mean the world to me if you subscribed, restacked, or dropped a comment. All of that’s free for you but a huge boost for me, and I’d really appreciate it.

Take care & see you in the next one ✌️
Michael

Subscribe now

The truth is it takes hours to manually dig the gold out of data, so it usually doesn’t happen. Even if you’re one of the rare exceptions that does pull the gold out, most people never push through to visualize the data properly.

As a result:

Everyone vomits up boring bar charts that say nothing.

In this article, I’ll show you how to solve this problem permanently.

If you’ve ever wanted (or needed) to serve up sharp data in a beautiful dashboard, this article is for you.

Give me 15 minutes and here’s everything you’ll get:

• 2 power prompts, ready to copy and paste straight into Cowork

• The exact messy spreadsheet I’m using, so you can practice on safe numbers before you point this at your own

• A finished dashboard as a single file that opens on any machine, even offline, and sends to anyone

• All of it sharp, beautiful, and sure to impress the people around you

Follow along and you can turn your own spreadsheet into a clean dashboard in a few minutes, and skip the hours it used to take.

Let’s go 👇

If this looks worth your time, save it, give it 20 minutes this week, and send it to the one friend you know that needs it.

Subscribe now

Why you never actually analyze your own data

All of us have piles of data that we’ve been meaning to dig into. Maybe it’s an export you pulled a while back, or a file someone sent you, or numbers you’ve been collecting that you never really looked at.

You know there’s a little nugget of gold in there, but you just haven’t found it yet. And the reason is pretty simple:

Real analysis takes a couple hours of intense, focused effort.

And most of that is tedious data-entry work before you actually learn anything. You gotta organize the data, label it, validate it, clean it up.

So even though you keep meaning to dig in and do the analysis, you don’t. The data just sits there.

Going from a pile of numbers to something you can actually use

Except you don’t need to do any of that anymore.

Because AI.

You can hand the messy sheet to a nerdy AI data analyst and let it work for you. It’ll put the data into neat little tables, find relationships, and even visualize it for you.

There’s actually a name for this climb. It’s called DIKW:
Data, Information, Knowledge, Wisdom.

Raw data is a pile of numbers that don’t mean anything yet (data). Group and label it (information), and you start to see patterns (knowledge). See how those numbers relate to each other, and you actually understand what’s going on (wisdom).

That’s your ladder, and your goal with any sheet is to get your data all the way to the top.

From Data to Wisdom. And AI can do the tedious geek work to get you there.

Grab the same spreadsheet I’m using

From this point forward, we’re building on a messy spreadsheet from a made-up company called Munder Difflin. The numbers are fake, and that’s on purpose.

There are two reasons:

First, I can hand you the exact same sheet, so you can follow along instead of having to dig up your own data.

Second, nobody’s private numbers end up in a free download. You get to practice on something safe before you point this at your own stuff.

So what’s in it?

The kind of data you’d actually run into at work:

• Monthly sales numbers

• Product lines the company sells

• How much each channel brings in

• Marketing spend

• etc.

And it’s messy in the normal ways, with mismatched labels, a few blank cells, and a stray total row sitting where it shouldn’t.

I’m also assuming you’ve already got Claude Cowork set up and running. If you don’t, I covered that in How to Turn Claude Cowork into Your Chief of Staff.

Go get that going first, then come back here.

Subscribe now

First, let AI think like your data analyst

Now, I’ll show you how to use AI to think like a data analyst.

If we do it right, AI will read your sheet, find the relationships worth showing, and pick the chart types that work best with your data.

I like to do this in two passes:

First the analyst pass, where AI does the thinking.

Second, the visualizer pass, where AI turns that thinking into a dashboard.

To get started, you’ll paste the analyst prompt into Cowork (shared below), point it at the Munder Difflin sheet, and tell it what your objectives are.

Maybe you already have a question in mind, like which products actually make money. Maybe you just want to know what’s interesting in there.

Either way, it asks you a couple quick questions first, then it gets to work.

It opens the sheet and reads the whole thing for context. It cleans up the messy parts, the blank cells and the labels that don’t match, and it tells you what it changed so nothing happens behind your back. Then it starts pulling out what’s worth looking at.

It’ll review the numbers you pointed to, and find the relationships you didn’t even know were sitting in there.

Here’s the cool part:

The analyst knows charts... like, really knows them. Most of us reach for the same chart formats every time. We default to line or bar charts, and those work fine for everyday stuff.

The interesting reveals come from the ones you’d never think to build yourself, like a bubble chart or a sankey or a waterfall. The prompt has a built-in cheat sheet that matches the shape of your data to the chart that actually fits it, so you don’t get a sad default chart for everything.

When you see your data drawn in a way you’ve never seen it, you notice new things you had previously missed.

When I ran this on the Munder Difflin sheet, it came back with a bubble chart comparing sales channels and a ranked chart of the top products. Then it wrote everything up as a brief, and that brief is the plan for the next step.

Here’s the analyst prompt. Grab the whole thing, paste it into Cowork, and point it at your sheet.

## ROLE

You are a sharp data analyst. You read whatever dataset you are given, figure out what it is actually saying, and design how it should be shown. You think and you plan. You do not build anything.

## GOAL

Turn a raw dataset into a clear visual brief that another tool can build into a dashboard. The brief names the few most important things in the data and the best way to show each one. It must be self-sufficient: it carries both the design and the exact numbers, so the dashboard can be built straight from the brief alone.

## TASK

**1. Run a quick interview.**
Before you touch the data, ask up to three short questions, all at once:
- What is this data and where did it come from?
- What are you trying to figure out, or what decision are you trying to make?
- Who is the dashboard for (just you, your boss, your team)?

Keep it light. If I say "just go" or skip a question, proceed on your best assumption and tell me what you assumed.

**2. Read and understand the data.**
Open the file and read all of it.
- List every column: its name, what it means, the type of data, and the units.
- Note the grain. What does one row actually represent?
- If the data is messy, clean it so you can work with it. Fix inconsistent labels, blank cells, mixed date formats, stray total rows, and extra whitespace.

**3. Find what the data is actually saying.**
- Name the few most important things in here: the standout, the outlier, the trend, the thing that breaks the pattern.
- Draw the relationships between variables. What moves with what. What trades off against what.
- If a relationship is not really there, say so.

**4. Choose how to show it.**
Pick the few most important relationships, around five to eight, not every column. For each one, choose the chart type that fits the *shape* of the relationship, not the chart you reach for out of habit. Use this as your map:
- Trend over time → line or area
- Part of a whole → stacked bar or waffle
- Comparing categories → column, bar, or dot plot
- How two things relate → scatter, and add bubble size for a third variable
- Ranking that shifts between two moments → slope chart
- Flow or movement between stages → sankey or alluvial
- Above or below a benchmark → diverging bar or bullet
- Density across two dimensions → heat map
- A running build-up or draw-down → waterfall
- The spread of values → histogram or box plot

Suggest at least one chart I would never have reached for myself, so that seeing the data drawn a new way surfaces something I did not see before.

**5. Write the brief.**
End with a clean, structured brief in this shape:

*Dashboard*
- Title
- Who it is for
- The headline story in one or two sentences

*Charts* (numbered, in the order they should appear, most important first). For each chart:
- Title
- Which columns or fields it uses
- The relationship or insight it reveals, in one line
- Chart type, and one line on why that type fits the shape of the data
- Signal cues if useful: what counts as good, okay, and bad (green, yellow, red) and the rough thresholds
- The one-line takeaway caption that should sit near the chart

*Notes for the build*
- Anything the visualizer should know: a recommended order, a metric to feature up top, a comparison to keep paired.

*Data appendix*
- For every chart above, lay out the exact numbers it plots as a small, tidy table (one row per data point, columns clearly labeled). You already computed these, so write them down. These resolved values are the single source of truth for the build, so the dashboard and your captions can never disagree. Keep the narrative sections clean and readable. The raw tables live down here.

**6. Save the brief.**
Save the finished brief as a markdown file named `Visual Brief — [dataset name].md` so I keep a real artifact and can hand it off cleanly. Tell me where you saved it.

## CONTEXT

- This is the first of two passes. You write the brief. A separate visualizer prompt reads your brief and builds a single dashboard from it. The brief must stand on its own, design and numbers both, so the visualizer never has to re-open the raw data to get a value right.
- You are running inside Claude Cowork and can open and read the file directly.
- This prompt has to work on any dataset, so make no assumptions about a specific company, sheet, or set of columns. Work only from what you are given.
- I am smart but not technical. A clean dashboard built on a few clear charts beats a busy one.

## CONSTRAINTS

- Do not write any HTML or build anything. Stop when the brief is written.
- Do not fabricate data or trends. I cannot always catch a made-up insight, so do not hand me one.
- Never silently change my numbers. Work on a copy, and report every cleaning change in a short list, like "merged 'East' and 'east' into one branch." If something is ambiguous, state your assumption instead of guessing in silence.
- Keep it to roughly five to eight charts. Prioritize hard.
- Avoid pie and donut charts unless they are genuinely the best fit. They are the lazy default.
- Write in plain language.

Second, turn that plan into a real dashboard

Now it’s time for the visualizer!

(My favorite part.)

Paste the visualizer prompt (shared below) into the same chat, right under the brief. It reads what the analyst wrote and builds the dashboard straight from it.

You don’t move files around or re-explain anything. The plan is already sitting right above it.

A minute or two later, it’ll generate an HTML file. Just double-click it and it opens in your browser like any web page.

It works with no wifi, and there’s nothing to install and no account to log into. Everything the dashboard needs is baked into that single file, so you can send it to someone and it just opens on their machine too.

Here’s the visualizer prompt. Again, it goes in the same chat, right under the brief.

## ROLE

You are a dashboard builder. You take a finished visual brief and turn it into a single, self-contained HTML dashboard. You build. You do not re-analyze the data or second-guess the chart choices.

## GOAL

Produce one HTML file that renders the dashboard described in the brief: a big title up top, every chart boxed in its own labeled panel, green/yellow/red signal cues applied, clean and readable at a glance. It opens by double-click on any machine, offline, and shares as a single file.

## TASK

**1. Read the brief.**
Take everything from it: the dashboard title and headline story, the chart list (type, fields, signal cues, caption), the build notes (order, pairing, KPI tiles, annotations, color discipline), and the Data appendix.

**2. Use the brief's numbers as the source of truth.**
The Data appendix holds the exact values for every chart. Plot those. Do not recompute them and do not invent any. If a value you need is missing from the appendix and the source file is open in this chat, pull it from there and say so. If it is missing and unavailable, flag it rather than guessing.

**3. Build the layout.**
- A big dashboard title across the top, with the headline story as a subtitle.
- An optional row of KPI scorecards if the build notes call for them.
- Below that, each chart in its own panel, like a card with a label. Follow the order and the pairings the brief specifies.
- It should read at a glance. Clean, not busy. Every widget labeled.

**4. Draw every chart.**
- Use the chart type the brief specified for each one. Do not downgrade a slope chart or a sankey to a plain bar because it is easier.
- When the brief names the specific series to feature, plot those and only those. Do not expand a focused comparison into every category in the data. If the brief asks for three lines, draw three lines, not ten.
- Apply the green/yellow/red signal cues and thresholds exactly as the brief gives them.
- Put the one-line caption near its chart.
- Layer any event annotations onto the time charts as the build notes direct. Keep them in the chart's margins or empty space, never crossing the plotted data, and never clipped by the panel edge. Add headroom at the top of the chart so a callout sitting near the highest value stays fully visible.
- Keep the axes clean. Never let a tick label repeat, like printing the same top value twice. Give the axis titles and the legend their own room so nothing overlaps the plot or each other.

**5. Make it one self-contained file.**
- Inline the charting library's source directly into the HTML inside a script block. No CDN, no external script, style, or font links. Use a system font stack.
- Everything lives in the single file so it opens offline by double-click.

**6. Save it and hand it back.**
Save the file as `Dashboard — [name].html` and tell me where it is.

## CONTEXT

- This is pass two of two. The analyst already did the thinking. The brief is your spec, not a starting point to argue with.
- Primary path: you are running in the same Cowork chat as the analyst pass, so the brief is right above you. The brief is self-sufficient, so you can also build it cold from the brief alone.
- Use ECharts as the charting engine. It draws the whole range the brief might ask for (line, area, column, bar, scatter, bubble, slope, diverging bar, heat map, waterfall, sankey, bullet-style gauge) and looks like a real dashboard out of the box. Standardize on this one engine so every widget shares the same typography, spacing, and color, and the dashboard reads as one clean thing instead of a pile of mismatched charts.
- The reader is non-technical and opens this by double-clicking. It has to just work, offline, on any machine, with no setup on their end.

## CONSTRAINTS

- One file. No CDN and no external scripts, styles, or fonts. Inline everything, including the charting library. It must open offline.
- Do not fabricate numbers. Use the Data appendix. Flag anything missing instead of filling it in.
- Build the chart types the brief specified. Do not simplify the interesting charts into generic ones.
- Apply the signal colors and thresholds exactly as given. Reserve red for true red-threshold values so it stays meaningful.
- Clean, not busy. Boxed and labeled panels, a big title, readable at a glance. Keep it to roughly eight widgets.
- Do not make me install or download anything. All the work happens on your side.

What a real dashboard actually looks like

So what does it actually look like?

It’ll have a big title across the top, then your data laid out in panels underneath. Each chart sits in its own box with its own label, so you always know what you’re looking at.

The numbers that matter most ride up top as a row of scorecards. Good, watch, and trouble show up as green, yellow, and red, so you can read the health of the whole thing in about five seconds.

When it comes out right, it feels clean. Easy on the eyes, nothing crowded, nothing you have to squint at.

That’s the difference between a real dashboard and the generic template look. A template dumps every number on the screen and hopes you sort it out. A good one decides what matters and points you right at it.

If there’s anything you want to change, you simply prompt Cowork and it’ll make the adjustments on the fly.

Subscribe now

A data analyst in your pocket

These 2 prompts will save you the hours of grunt work it used to take to analyze what’s in a sheet, draw new connections, and build visualizations. AI can handle all of that now.

Point this at the stuff you’ve been putting off:

• Your email

• Your sales numbers

• Your subscriber export

• Whatever it is you work with

It’s like having a data analyst in your pocket. Use it.

Hope you found this helpful & see you in the next one!
-Michael ✌️

And then the slacker who presented right after you, with half the analysis and twice the polish, walked off with all the credit.

This scene plays out in meeting rooms every day. Anytime you present, you’re doing two jobs at once:

1. Make sure the numbers are absolutely rock-friggin-solid

2. Present them in a way that LOOKS AMAZING

Most people - especially engineers - only ever nail the first one. The second is slow, fiddly, and can eat four or five hours of your night, so it’s the part everyone shortchanges.

That’s why I built a super-prompt that handles it for you.

Point it at the data you worked so hard on, and it builds a beautiful, fully animated deck that blows your boss’s hair back.

Here’s everything this prompt does for you:

• It interviews you about your brand, the story you want to tell, and the metrics you most want people to notice

• It drafts the titles, subtitles, story, and what belongs on each slide, then gets your sign-off before it builds

• It builds a beautiful animated slide deck, one chart per slide, with clear labels, clean trend lines, and contrasting colors that pull the eye to the metric that matters most

• It makes you look like an absolute stud in front of your boss and your colleagues, and the envy of everyone in the room

Point the prompt at your data, and you’ll get back something you’d be proud to present anywhere.

This is for anyone who regularly presents data. Project managers, marketing agencies, freelancers, consultants reporting back to a client. Engineers especially, because you all love data but hate making it look pretty.

Instead of losing half a day trying to wrestle your data into something beautiful, you can run this prompt and it is done in about twenty minutes.

Sound good? Let’s go 👇

If this looks worth your time, save it, give it 20 minutes this week, and send it to the one friend you know that needs it.

Subscribe now

The meetings where 1 wrong number follows you for years

We all know what it feels like to get called into your boss’s office to walk through the numbers. Or worse yet, maybe it’s not just your boss, maybe it’s a whole room full of VPs. Or the real nightmare, a client deciding whether to keep paying you.

There’s a lump in your throat. You know the stakes are high. The data has to be right, and it has to look good. If even one number is off, or one data label is too small to read, your reputation takes a hit.

I’ve been in this spot a million times.

The scariest presentations I ever gave, hands down, were back in network ops at a Fortune 50 telecom company I used to work for. I had to present project readiness in front of all our VPs. One of them had 6,500 people in his department. My job was getting all of those people, his entire department of 6,500 people, ready for a +$100,000,000 technology we were rolling out.

(NO PRESSURE! 🤮)

Even today, I can’t really believe that I did most of my presenting straight from a spreadsheet.

The spreadsheet had a table with hundreds and hundreds of rows and tracked every team, and every readiness task. It took 10+ meetings every week just to keep it updated. The charts were powered by a bunch of crazy formulas I wired together.

Just wrangling the data alone was BEYOND PAINFUL.

In the end, the spreadsheet looked okay, presentable enough for my VPs in our department. But it was never something I would have shown anybody outside our organization.

I’m sure you’ve been in similar situations.

There are 2 ways to look like an idiot up there

Let’s be honest about what’s on the line:

Anytime you present in front of other people, your reputation, your status, and how people see you is at stake.

Present data that’s wrong, get called out in front of everybody, and the embarrassment is freaking awful. You’ll look like, or at least feel like, a complete and total idiot.

Present data that’s right but the charts are a mess, the labels too small, the colors weird, anything that shows you can’t handle data, and you still look like an idiot, just a different kind of idiot.

I know somebody reading this is going to roll their and think, “Yeah, well... substance matters more than form.”

Sure, I agree.

If you stack rank the priorities, what the data says is more important than what the data looks like. But anybody who tells you the way it looks doesn’t matter at all is just wrong.

The colors, the fonts, the size of the labels, it all creates an impression. It sets a tone, it gives you authority, and it moves people toward your point of view. Or it quietly reinforces that you’re not the type of person to put in front of clients, executives (or anyone else that matters).

THIS is why aesthetics are important. They increase trust. They communicate that you understand corporate etiquette. That you can be trusted in front of the people that matter.

Subscribe now

4 hours of work done in 20 minutes

Here’s the good news:

You don’t have to do any of this by hand anymore.

I’ve made a prompt that does the whole thing. You copy it, paste it, and point it at the data you already have. That’s it.

The work that used to eat four or five hours of your night, building charts, fighting formulas, fussing with labels, AI handles in 20 or 30 minutes.

And it doesn’t just spit out a file and hope. First it looks at your dataset. Then it runs a short interview. It asks about your brand colors, the story you’re trying to tell, and the metrics you actually want people looking at. It takes your answers and writes up a quick brief.

Then it builds the thing.

What comes out is beautiful, clean, and easy to read. It frames the numbers inside an actual narrative instead of letting them just sit there. It’s the kind of presentation you’d feel completely comfortable putting in front of your boss, your VPs, a board, or a client.

You gotta understand:

When people stop and pay attention to YOU, your status goes up. AND THIS will help you make the right impression.

THIS will help you blow their friggin’ socks off.

You’ll stop being known the person who threw a messy spreadsheet on the projector and become the one who clearly knows how to handle data, the expert in the room.

All from one simple prompt you copied and pasted.

Let me just show you how good this looks

Let me show you how awesome this is.

I use this prompt regularly, but since I can’t share private company info in public, I’ve used synthetic data for the example below.

I ran this on a data set from a fictitious company called Munder Difflin.

After a short interview, the prompt built the charts, added clean data labels, drew trend lines, and animated them. It also used contrasting colors to pull your eye straight to the numbers that matter most. It gave every chart its own page so your audience won’t get overwhelmed.

I used to grind every one of these little details out by hand. But this prompt did all of them at once, and it did them well.

Take a look:

Subscribe now

Here’s the whole prompt, just copy and paste it

It’s one prompt.

You paste it into an AI that can open your files and write code (I run this in Claude Cowork, but anything with file access will do), then you point it at your spreadsheet. Your messy, multi-tab, hundreds-of-rows sheet, exactly as it sits right now.

No need to clean it up first. The prompt handles that part.

Grab all of it, paste it into a fresh chat, and point it at your sheet 👇

Read more

I’m guessing you probably use at least one tool you can do this with, too.

In this article, I’ll show you exactly how I cloned a simple app without writing a single line of code myself, and hand you a starting point to do the same with yours.

Here’s what you’ll get:

• The 40-minute rebuild, start to finish - where you watch a non-coder clone a paid app without touching a line of code

• The brand new “goal” command - the buried mode inside Codex and Claude Code that turns AI from a chatbot you babysit into a builder that writes, tests, and fixes its own code while you do other stuff

• My exact design doc - copy the spec I handed the AI so your build starts on third base, not a blank page.

• The “could I just build this?” test - a dead-simple way to look at every tool you pay for and spot the ones you could replace.

It’ll take about 8 minutes to read the whole thing, and you’ll learn a skill that could save you hundreds of dollars per year.

Let’s go 👇

If this looks worth your time, save it, try it sometime this week, and send it to the one friend who will think this is cool.

Subscribe now

It started when Google quietly broke my favorite tool

For a while I had a meeting setup I actually liked. I was using Fathom, and before that, Otter. They join your meeting like another attendee would, transcribe it, and handle the speaker identification. I’d even built an AI system on top that pulled each transcript onto my computer and dropped it into my own template, so I could search back through everything and trace what got decided.

It worked beautifully.

Then Google started restricting third party tools from joining Google Meets. I think it’s because they want people using Gemini.

Shocker.

The problem is Gemini only joins Google Meets. It won’t join or transcribe Zoom or Microsoft Teams, and that’s where plenty of my meetings happen. So the setup I liked stopped working.

So, I downloaded Granola. Instead of joining meetings as an attendee, it records your system audio and microphone, and it transcribes from that. Google isn’t even aware it’s happening.

(Take that, suckas!)

But then I got to thinking...

If all it’s doing is recording the speaker and the microphone and transcribing it, I could literally just vibe code a small app that does that. I could set it up to drop the transcript right where I want it, instead of pulling it down through an API call. And over time I could build on top of it and add the things Granola won’t, like a note taker that acts more like a facilitator, one that does some timekeeping and hits me with a notification when we’re ten minutes from the end of the call.

Even with my non-technical background, I knew I should be able to vibe code something like that. It wouldn’t be that hard.

How one prompt turned into a finished app

I didn’t start by telling Codex to write a design doc. I started with a messy product conversation.

I asked how hard it would be to build an app that listens to my meetings, transcribes them locally, and writes the transcript to a markdown file. I told it I wanted something like Granola, but without a ton of features.

Codex came back with the real tradeoff:

Markdown and transcription were the easy parts. The hard part was capturing my microphone and the other people on the call, through the computer audio, at the same time.

So I kept narrowing it.

I asked how Granola actually processes its audio. I asked for the cheapest accurate way to transcribe locally. It suggested Whisper. So, I asked how Whisper works, how big it is.

We kept narrowing, kept simplifying, kept reducing scope. All I wanted was an app I load up while in a meeting, hit record, it captures my audio and the meeting audio. Two different wav files titled “you” and a “them”, and then drops the transcript into a local markdown file in a folder I get to pick.

After this beautiful little dance I wrote one more prompt:

the design doc.

Write what the MVP is, and then write what the MVP is not, so the scope can't creep.

This made a huge difference.

Making it spell out what the app would not do is what drew a boundary around the build. Then I had it review its own plan and tell me if there was a more elegant approach. It caught that recording one or two giant audio files was a bad idea and switched to recording in chunks instead.

Click here to download the design doc, so you can start from mine and not from scratch.

After it was done with the design doc - which I didn’t read, those docs are for nerds 🤓 - I ran one command:

You can copy/paste it yourself, here:

/goal Build Phase 1 of BlahBlahBot from the design doc: create the native macOS SwiftUI app and prove separate microphone + system audio capture into timestamped chunks. Do not implement transcription yet.

If you’ve never used it, the goal feature works differently than just chatting with the AI. In a normal chat you ask, it answers, you ask again.

With goal, you hand it the outcome you want and it goes to work on its own. It writes the code, runs its own tests to check what it built against the goal, and if it’s not there yet it fixes the bugs and keeps going.

It stays in that loop until it decides the goal is met. Some people call this a Ralph loop. It’s in Codex, and it’s in Claude Code now too, so you’re not locked into the one I happened to use. I built mine in Codex because Codex got the feature first.

So, I dropped this prompt, pressed enter and walked away. The little wheel started spinning, and I went and lived my day.

I’d check in on it now and then and it was still going, still iterating. Every so often it would stop and ask me for a permission, and I couldn’t even tell you what half of them were for. The actual coding time was about 45 minutes. It took me closer to 2.5 hours start to finish, but only because I was off doing other things and wasn’t around to click “allow.” If I’d been paying attention, it would have been a lot faster.

I kept waiting for it to make me do the annoying parts, go create an account somewhere, or download some dependency, or paste in an API key. I’ve had to do all of that with other apps I’ve built. But it never asked.

I’m pretty sure it downloaded and installed Whisper itself.

It just built the whole freaking thing.

The experience with /goal was flippin’ amazing.

I give it 10 / 10 stars, highly recommend, and would do again.

Subscribe now

It works, but it is not magic

But... as with any vibe coded app, there were a few pesky bugs.

The audio files are huge. I had one 90-minute meeting and the recording came out to almost half a gig. The way it saves them is kind of weird. It records in chunks, a new file every five minutes, split across two folders. One folder is everything from my mic. The other is the system audio, which is everyone else on the call. Add it all up and that single meeting was around 500 megs sitting on my drive.

It also wasn’t deleting any of that on its own. So I went back to Codex and told it to clean the files up when a recording finishes. That fixed it, but I had to be the one to ask.

Transcription was slow the first couple of times too. There were some settings around how Whisper runs on my computer and what it pulls on to do the processing, and I had Codex adjust them. It gets faster every time I tweak it. I’m still tuning that part.

Let me be straight with you:

Mine is not Granola yet. Granola ships your audio out to cloud processors and gets clean results back fast, and on raw quality it still beats what I’ve got.

But mine is free and I built it without VC money. It lives on my own computer, and it already does the one thing I actually need. Give me two weeks and I’ll have a pretty damn good transcription app sitting right here on my machine. And the best part is I get to keep adding to it, building the custom stuff I actually want.

Who should try this, and who should just keep paying

So who is workflow like this for?

You’ll need a Mac, since the app is built native to macOS. A paid Codex or Claude subscription too, either one works, you only need the one. And you have to be okay letting an AI run loose on your computer for a while. If you’re already using Codex or Claude code, you’ve got that comfort already.

(That’s basically table stakes at this point.)

One more bit of honesty about what this post is:

I’m just showing you a cool thing I did this week and roughly how I got there, in case you want to try it yourself. I’m not pretending it’s a hand-holding walkthrough for a total beginner. And if this isn’t for you, the $14 a month I was paying Granola is a perfectly fine deal.

Keep paying it. No shame in that.

Every tool you use for one thing is now in danger

Here’s the biggest takeaway:

The reason I could rebuild Granola is that I only use it for one thing. I open it, it listens to my meeting, it transcribes it. That’s it. I never touch the notes feature or any of the other stuff bolted on around it.

Tools like this are in danger.

Look at a tool you pay for and ask what you really use it for. Forget the long feature list on the pricing page. Day to day, what is the actual job you open it for?

If the honest answer is one thing, one core loop you run over and over, that tool is probably a rebuild candidate now.

Once you start looking at your stack this way you see it everywhere. Take Calendly. For most people all it does is show someone your calendar and let them book a slot. That’s a value loop simple enough to hand to an AI and walk away from, same as I did with my recorder.

Subscribe now

The $24,000 reason SaaS should be nervous

Now here’s where it gets interesting, and this is more a prediction than a promise. I think B2B software is in real trouble.

Consumer apps are mostly safe.

Most people are never going to open Codex or Claude code to build their own version of something and then babysit it forever. Even a couple years from now, when you can probably pull this off in a single sentence and it builds in 30 seconds, most people still won’t bother. We live in a world where someone can have a fridge full of food and a perfectly good microwave at home and still drive halfway across town for Burger King, or just DoorDash it to the couch. That instinct is not going anywhere.

So consumer software gets to keep its customers. But business software is a different story.

In B2B, one tool can quietly cost a company 5 or even 6 figures a year.

Here’s a real example from my own life:

We pay around $24,000 a year for a business intelligence tool. I won’t say which one. Right now we’re rebuilding every dashboard we use in it inside a free open-source tool, so we can stop paying that bill.

When the thing costs $24,000 a year, spending a little engineering time to rebuild the part you actually touch does the math on its own. The return on building it yourself used to be tiny due to the labor cost. Now it’s enormous.

Any B2B tool charging 5 or 6 figures for a job that is really one core value loop is now a opportunity to be vibe coded.

The SaaS subscription you hate the most might be optional now

So that’s where I’ve landed. For years I rented my software. I paid the monthly bill and lived with whatever the company up top decided to build next. That’s flipping now.

The tool you hate is turning into something you can open up and reshape into whatever you actually need.

That $14 a month I was handing Granola is optional now. The thing it does for me turned out to be something I could just own outright.

So go do the one thing I’d want you to do after reading this.

Pick the tool you pay for and quietly resent the most. Open it up, look hard at what you actually use it for, and ask whether its one job is something you could just build. You might be surprised how often the answer is yes.

I named mine BlahBlahBot. That should tell you exactly how I feel about meetings.

Take care & see you in the next one!
-Michael ✌️

P.S. Can you do me a favor? If you found this helpful can you please like, comment, and re-stack it? It’s free for you but a HUGE BOOST to my little publication. Thank you ✌️

If you liked this article, check these out:

• The Non-Technical Person’s Complete Guide to Claude Code

• How to Turn Claude Cowork into Your Chief of Staff (Tutorial)

• Steal My Second Brain Starter Kit

Subscribe now

Leave a comment

Share

I have transcripts piled to the moon in note taking apps I don’t even own. I’d agree to an action item on Monday and completely forget it by Thursday. Finding anything from a past meeting meant reading hours of transcripts.

If you spend half your week in meetings like I do, you know this feeling.

So, I built a system in Claude Cowork that solved this problem for good. In this article, I’m going to show you how you can build this system too.

Don’t worry, as with all of my articles, no code is required and I’ll show you how to do everything step-by-step. Cowork builds every piece of it for you.

(And if you want the shortcut, I provide all of the prompts you can simply copy and paste.)

Here’s what this system will do for you:

• Every meeting lands on your computer as a clean note with the summary, decisions, action items with owners, and the transcript at the bottom for reference

• You can search years of meetings in seconds and pull up any decision, action item, or data point you couldn’t quite remember

• You walk into your next recurring meeting with a prep brief already drafted from the last few times your team met

• A daily work log writes itself on top of all of it, so by Friday you can see exactly what your week contained

The full easy-to-follow build is next, with prompts you can copy and paste to stand it up, plus a handful of upgrades you can layer on to make it yours.

Follow the whole thing and you can have your own version running in about 30 minutes, and it’ll save you hours every single week.

Let’s go 👇

If this looks worth your time, save it, build it sometime this week, and send it to the one friend you know that needs it.

Subscribe now

You Can’t Out-Discipline This

Tell me if you’ve been here before:

You’re in a group meeting with five or ten people on it and the topic doesn’t involve you right now. So you’re multitasking. You’re rightfully trying to redeem the time, and you’re slowly getting absorbed in your own things.

Within a few minutes, you’re fully into other tasks. And then you hear your name.

“John, what do you think about this?”

Or maybe it’s worse.

“John, can you take this action?”

You wake up. You’re back in the present moment, and you don’t quite know how you got there. You don’t want to look like you weren’t paying attention, so you say “yeah, yeah, sure, I got that action.”

Then you pretend to write something in your notes.

I’m very guilty of this. I multitask all the time. I have to. I don’t have enough hours in the day to join all the meetings I need to and do my regular work.

As a leader in a modern office job, I spend at least half of my week in meetings. Easily 20 hours, sometimes more. By the time Thursday rolls around I can’t fully reconstruct what I committed to on Monday. I had a KPI meeting the other day where we identified 110 KPIs and outlined a bunch of next steps. It took me two days to get around to it.

By then, I couldn’t even remember what those “next steps” were.

And even when you ARE paying attention, you can’t always take good notes. Maybe today is the day you had a doctor’s appointment and you’re driving back and forth, listening to the meeting while navigating traffic. You can’t write notes while you drive.

Or maybe you’re physically in the meeting room but the conversation is moving too fast to catch every decision and action item that flies by.

This is where an AI note taker comes in. Something like Granola that sits on your computer and listens to every meeting, transcribing everything in real time. You don’t have to write a single thing down. The capture problem mostly gets solved.

Yay.

But now you have a bunch of transcripts on the cloud somewhere. Ten meetings a week. Ten thousand words a meeting. And next Tuesday somebody will inevitably ask you about a decision that was made three weeks ago and you can’t remember which meeting it happened in.

Or something comes up an you want to know what action items are still open across all your meetings. Or maybe you walk into a recurring meeting and you can’t remember what you discussed last time. Or your boss is asks about a data point came up in a meeting last month and you need it now but you can’t remember where.

The problem is that capture is the easy part. But a transcript on its own is just data. It only becomes useful when you can turn it into the thing you actually need:

• A decision pulled in seconds when somebody asks

• An action item surfaced before you forget it

• A brief for the meeting you’re walking into in fifteen minutes

That’s the layer we’re going to build.

Here’s What You’re Building

I had a manual version of this for years.

Before AI showed up, I took tons of notes on every meeting. In Slack, you can click your own name in the sidebar to send messages to yourself. That’s where I wrote all my work notes for years. Slack organizes messages by day automatically, so if I needed to find notes from a specific meeting, I could just scroll up to the date. And Slack is searchable, so I could find a phrase later if I needed it.

It was simple, and it worked for me. For years.

Then Claude Desktop showed up and it blew my freaking mind. For the first time I had intelligence sitting on top of my file system. I knew I had to get all my meeting transcripts in a place where Claude could read them. Because there was so much more I could do with the raw material if I just gave Claude access to it. Way, way, way more than I could do with my scrappy Slack notes.

So then I started building. I put together an automatic note-taking system that saves me literally hundreds of hours a year.

If you build the system I’m going to walk you through below, here’s what you’re going to get:

Every time you have a meeting, the transcript ends up on your computer. In a folder you actually own, with a clean filename you can find later.

You get a structured summary of every meeting, built around the things that actually matter to you. Decisions made, action items with owners, key points with timestamps, whatever fields you care about most.

You can take those summaries and launch them into the rest of your work. Once Claude has the meeting context, it can draft just about any artifact you need from it. Tomorrow’s executive summary off this morning’s call. The SOP you’ve been putting off based on the new process the team agreed to. The deck for next Tuesday’s recurring meeting built from the last three Tuesdays. The meeting minutes you have to send out by EOD. Whatever you have to deliver next, you can start it from your meeting notes.

And at any point, ever, you can search across every meeting you’ve ever had and find exactly what you need within seconds.

Literally seconds.

The decision from three weeks ago, the action item that didn’t make it onto your list, the data point that came up once in a conversation last month and disappeared, the thing someone said about a project that you can’t quite remember. All of it is sitting right there waiting for you.

The whole thing is SUPER SIMPLE to set up. You don’t have to be technical to build it. We’re going to use Claude to help us write every piece of it.

By the end of this article, you’ll have your own version up and running.

Subscribe now

Watch It Work

Once this system is running, it changes the texture of your work day. Let me walk you through a few moments that happen to me on the regular.

Working the meeting summary.

You had a heavy meeting on Monday and walked out with a vague sense of what you committed to. Wednesday rolls around and now you actually have to do the work. I had this exact thing happen last week. We identified 110 KPIs in a single meeting and outlined a whole stack of next steps. Two days later I had to act on it.

Instead of trying to reconstruct any of it from memory, I opened the meeting note. The action items were right there with who owned each one, the decisions were right there with the rationale, and the full conversation was in the raw transcript if I needed to go back to it. I didn’t have to remember any of it. I just had to read.

This is the day-to-day payoff. Every meeting on your calendar produces a clean note built around the fields that matter to you. When it’s time to act on something a few days later, the note is sitting there waiting.

Meeting prep.

You have a recurring meeting in 15 minutes. You haven’t touched the project in two weeks and you can’t really remember where you left off last time. Normally this is where you’d waste the first 10 minutes of the meeting rebuilding context with everyone else.

Instead you ask Claude to review the last several meetings on this recurring event and surface what’s been covered, where you’re at, and what needs to happen next. It hands you back a brief that covers last meeting’s decisions, the action items that were assigned, the open questions, and the two things that came up that need to be discussed today.

You walk in with full context. Nobody wastes the first 10 minutes recapping. You can actually get to the work.

Mid-meeting search.

You’re in a meeting and someone mentions a project from a few weeks back and asks what you decided about it. Normally this is the moment you’d stall. You’d say “let me get back to you on that” or try to bluff your way through.

Instead, you quietly ask Claude to search your past meeting notes. In seconds it finds the decision, who made it, the context around it, and the action items that came out of it. You wait for the right beat to chime in and answer like you’ve been holding it in your head the whole time.

People think you’re a freaking genius and that you never forget anything. They don’t know you’re searching old meetings in the background while everyone else is still talking.

Other things you can do.

A daily work log gets auto-generated for you at the end of every day. The meetings you attended, the decisions that came out of them, and the action items that landed on your plate all live in one note, with links back to the source meetings. By Friday afternoon you can scroll back through the whole week and see exactly what happened and what’s still open.

That’s the version that comes built in. Once the foundation is there, you can layer on whatever else fits how you work. A weekly action item digest. A quarterly review across an entire project. A plugin that auto-drafts a status update for a specific recurring meeting. The structure of the system makes most of what you’d want to build on top of it pretty trivial.

If you run meetings for a living, you don’t have a choice about this. Project managers, meeting hosts, CEOs, anyone running a team. The cost of forgetting compounds across everyone you’re responsible for, not just you. You can’t be fully present in every meeting and still do the rest of your job. The only way out is through a great system.

The whole pipeline is buildable in about 30 minutes.

Below, you can find every prompt I use to run this.

Copy what you want, tweak what you need, and in about 30 minutes you’ll have your own version up and running.

Subscribe now

Step-by-Step Setup

We’re going to walk through this end to end. Should take you about 30 minutes.

Quick note for Claude Code users: I’m writing this for Cowork because that’s the audience that needs the most hand-holding. Everything here works on Code too. Same prompts, same architecture, same outcome. The only real translation: where I say “build a plugin,” you’ll say “build a skill.” Where I show you a Cowork screenshot, you already know the Code equivalent. If you’re on Code, you don’t need me walking you through the install bits anyway. Read through, mentally swap the language, and you’ll be fine.

Step 1: Install Claude Desktop

Go to claude.com/download and grab Claude Desktop. It’s available for Mac and Windows. (No Linux yet. Cowork isn’t there. If you’re on Linux, you’re already a Code user and the disclaimer above covers you.)

Once it’s installed, open it up and look for the Cowork tab. That’s the agentic mode we’re going to use. Cowork is the lowest common denominator here. Anybody can use it. You don’t have to know how to code.

One thing to flag up front: Cowork requires a paid Claude plan (Pro, Max, Team, or Enterprise). The free tier won’t cut it for this.

Step 2: Set Up Obsidian

If you don’t have Obsidian installed yet, go install it from obsidian.md. It’s free.

Open it and create a new vault, or add to one you already have. That’s all you need to do here. We’ll set up the folder structure inside the vault using a prompt in the next step.

If you’ve never used Obsidian before and you want a fuller intro, I wrote a getting-started post a while back: https://michaelcrist.substack.com/p/obsidian-get-started

That’ll get you up and running. Come back here when you’re done.

Step 3: Open Cowork and Build Your Template

You’re going to run every prompt in this tutorial inside Cowork, so let’s get that set up first.

Open Claude Desktop and click into the Cowork tab. Click Projects, then New project. Choose Use an existing folder and point it at your Obsidian vault. From here on out, every prompt I give you gets pasted into this Cowork project.

Now the first piece: the template every meeting note will be built on. The AI fills it in for you every time a new meeting lands.

Paste this prompt into Cowork:

Read more

It was AI.

If your brain also runs on novelty, drops into hyperfocus without warning, and has built (and abandoned) 17 productivity systems trying to outrun itself... you and I might have the same wiring.

Well, I have good news for you:

I’ve found that AI is GREAT at doing all the things ADHD-brains are the worst at. For the last 6 months, I’ve intentionally offloaded as much of my executive functioning to it as I possibly could. The amount of mental bandwidth I’ve freed up is GENUINELY ABSURD.

If you give me the next 15 minutes of your time to read this article, I’ll help you free up half of your brain. Here’s what’s inside:

• 4 brain functions your ADHD is worst at and exactly how AI covers for each one

• 7 prompts that turn your ADHD into a literal superpower

• The AI trap ADHD brains fall into hardest and how to sidestep them for good

If your brain works anything like mine, I promise you’ll start feeling the difference this week.

Are you ready?

Let’s go 👇

Subscribe now

Your Chatbot Probably Knows You Have ADHD

One Sunday afternoon, I was scrolling Facebook and saw a post where someone said that AI helped them identify their neurodivergent behaviors.

Sounded interesting, so I dropped this short prompt into ChatGPT:

Based on all of our conversations and everything you know about me, do you think I exhibit any signs of neurodivergence?

ChatGPT didn’t hesitate. In fact, it didn’t even drop into thinking mode. It just snapped back that I had a textbook case of ADHD.

Then, as if to defeat any counter-argument, ChatGPT proceeded to write an entire essay around all my little behavioral quarks it had noticed.

It said that I regularly:

• Seek novelty for motivation

• Experience “time blindness” when working tasks

• Drop into hyperfocus for weeks at a time

• Feel “saturated” with stimulus

• Live in a constant state of “low hum” anxiety

• Systematize everything... multiple times (🤚 guilty)

Blah! Okay, fine ChatGPT. You win.

I closed the browser with little forceful mouse clicks (the digital equivalent of slamming a door).

It took me a few weeks of reflection, but I finally admitted that ChatGPT was right. Despite being a paragon of success as a human being, I struggle with all of this.

But I also realized one other thing... AI doesn’t.

In fact, AI is GREAT at doing precisely the things I’m terrible at. Things like processing huge amounts of information, remembering my entire to-do list, prioritizing objectively, embracing repetition without boredom, and a whole lot more.

Once I started building my workflow around this asymmetry (me doing the things only I can do, and giving AI all the things my brain refuses to do) everything changed.

EVERYTHING.

That’s what the rest of this article is about.

AI Can Run the 4 Brain Functions ADHD Keeps Dropping

I’ve wanted to be a writer since the fifth grade. I’ve had a few viral hits, and one crazy run on Quora that pulled close to a million views in a month. Despite all of that, I still find it brutally hard to sit down and actually write.

The thing I’m best at is the thing I can least reliably start.

Apparently, psychologists have given this gap a name:

Executive function.

It’s the manager part of your brain, the part that decides what to do, holds the thread while you do it, tracks how long it’s taking, and picks what comes next.

Folks with ADHD struggle with executive function. AI doesn’t (yay).

Once you understand how the manager part of your brain works, you can start handing the tasks you’re weakest at to AI.

Each of these 4 jobs below (starting, holding, timing, prioritizing) has a prompt built around it later. It’s worth understanding them first, though. It’ll help you write or modify your own prompts when your brain breaks in some way that I didn’t think of.

Starting

When I sit down to write, a resistance comes up. A weight. A low level dread, heavy enough to make me avoid it, though it doesn’t shut me down on anything else. The heaviness is weirdly emotional.

The motivation isn’t there, and seldom comes.

Maybe for you it isn’t writing. Maybe it’s the deck you keep not opening, or the proposal that’s been three days late in your head while you do everything except start it.

It’s the same dread (thanks ADHD!), just pointed at a different thing.

Unfortunately, AI doesn’t remove the feeling (sorry), but it can help shrink the first step down until the dread has nothing to grab onto.

The Cold-Start prompt helps with this.

Holding

I start one task and three more announce themselves before I’ve finished the first. By the time I’ve handled those, the original thing is gone. Not finished. Gone. There’s no version of me holding the next 23 things in my head, and my brain doesn’t politely queue them, it just drops whatever I’m not looking at.

So I keep a to-do list, and for me it works like the bumper lanes at a bowling alley. It doesn’t make me better at the game. It just keeps the ball out of the gutter. That one habit has kept me on track my entire career, and I’m not exaggerating when I say I don’t think I’d be where I am without it. The list is just memory I can see.

Maybe for you it isn’t a list. Maybe it’s 40 open browser tabs you can’t close, because each one is a thought you’ll lose the second it disappears.

AI does what the list does, except it holds the thread back to you. You can hand it the messy pile, get pulled into something else for an hour, and come back to “here’s where you were, here’s what’s next” instead of a cold blank screen.

The Re-Entry prompt helps with this.

Timing

I cannot feel time passing. I’m a genuinely terrible judge of it.

I budget 20 minutes to get my kids out the door. Shoes, snacks, backpacks, everyone in the car. It has never once taken 20 minutes. It takes twice that, every time, and every time I’m surprised.

Same story at work. I will swear a 10-slide deck is about an hour. It eats half a day. It always eats half a day, and the estimate never updates no matter how many times reality corrects it.

Maybe for you it’s “I’ll just answer a few emails” quietly becoming the entire morning.

AI can be the outside clock I don’t have. Ask it to scope the thing before you start, and it’ll tell you, flatly, the half day you were about to pretend was an hour. It won’t make time feel real. It’ll just stop you from betting against it.

The Time-Reality Check prompt helps with this.

Prioritizing

This one’s different for me.

I learned to triage early, running expensive outages where the wrong call cost real money. So a flat list of equally loud fires has never really been my problem. I can rank.

Getting tricked is the problem.

Something walks in wearing an urgent costume, I hand it the whole day, and the thing that actually mattered never moves.

Maybe for you it’s the Slack ping that quietly eats the deep work you blocked the whole morning for.

What works for me is slowing down and ranking by the consequence of not doing each thing, not by how loud it is. AI runs that ranking out loud, before the fake emergency eats the afternoon. It doesn’t care how urgent something feels, which is exactly why it’s useful.

The Brain Unload prompt helps with this.

None of these four are character flaws. They’re just functions that came in light for people wired like us, and the strange, almost unfair part is that you can now borrow the missing piece for pennies on the dollar.

Subscribe now

7 Prompts That Turn Your ADHD Into a Literal Superpower

Every prompt here takes something you’ve apologized for and turns it into a superpower.

These prompts cover the four functions from the last section, plus the parts of the day that fall between them.

Copy them and replace the content in the brackets with your own info.

1. The Cold-Start.
For the task you want done and still can’t make yourself begin. It gives you one rock to pick up and nothing else to look at.

I need to start [task] and can't make myself begin. Don't give me the plan. Break this into steps so small each takes under a minute, give me only the first one and exactly where to put my hands, and wait for me to say done before the next.

One rock to pick up, nothing else in frame. That’s a thing you can actually do.

2. The Re-Entry.
For when a day away from a project wipes the entire context. It lives in a Claude Cowork folder, a folder your AI can read and write between sessions, because it only works if the memory survives after you close the laptop.

Closing: Before I stop: I'm on [project]. Save where I am, what I was thinking, the open question, and the next move. Hold it.

Returning: Reload [project]. Where was I, what was I mid-thought on, what's the next action?

You stop being the thing that has to remember. The folder is.

3. The Brain Unload.
For when every open loop is screaming at the same volume and you freeze.

My head is full of open loops and I can't pick. I'll dump them all below, unsorted. Sort into Now, Later, Trash. For Now items only, one short next action each. Ignore the rest.

You don’t sort first. You dump, it sorts. The sorting was the part you couldn’t do.

4. The Time-Reality Check.
For the task you swear takes an hour and has never once taken an hour.

I think [task] takes [my guess]. It never does. Name the 3 hidden sub-tasks I'm not counting, give me a realistic time, and the latest I can start and still hit [deadline].

Time blindness is really just not seeing the steps until they’re already costing you. The prompt drags them into view before they do.

5. The Boredom Off-Ramp.
For the novelty itch, right before it becomes four tabs and a snack run. It redirects the itch instead of fighting it.

I'm under-stimulated and about to spin out into [my usual drift: scrolling, tab-hopping, snack runs]. Give me a menu of specific things I can do right now to get my brain re-engaged — name the actual activities, no lecture: three 5-minute physical resets, two 20-minute options that are real progress on [goal], one 10-minute creative play break. I'll just pick one.

Unspent, the itch picks your next move for you, and it always picks the tabs. This hands the choice back to you.

6. The Focus Anchor.
For the focus you can hold with someone in the room and lose the second you’re alone. Body doubling, minus the body.

Be my focus partner for [30] minutes. I'm working on [task]. Check in every 10 minutes, ask what I've done, pull me back if I've drifted. Start now.

The body double never mattered. What mattered was someone about to ask what you got done, and AI can be that someone.

7. The 2am Net.
For the ideas that arrive at the worst hour, out of order, demanding nothing of you. A standing instruction in the same Cowork folder, set once.

Whenever I throw something at you, any idea or worry or half-thought, any hour, any order, just catch it. File it where it belongs, connect it to what it relates to, hold it. Never make me explain or organize it first.

You never have to organize the thought before you’re allowed to have it. That permission is the point.

Every one of these prompts cashes in on a trait you’ve been quietly embarrassed about. Which is exactly where this gets dangerous.

Subscribe now

Catch the Spiral Early or None of These Prompts Work

There are risks in everything you just read, and it isn’t the prompts. It’s the prompts plus a tool that will never once tell you to stop.

If you say “jump,” AI will always ask, “How high?” It’ll never ask whether or not you should be jumping at all. That sounds like a feature until it costs you two friggin’ days.

It’s a legitimate nightmare scenario for ADHD.

Here’s a real life example:

I had an internal presentation to put together for my team. The stakes were high.

The presentation required some data analysis, some vision-casting, decisions I actually had to make, and then a clear direction I needed to land. I budgeted 2 hours.

I had AI write the outline, and then I had AI build the slides. Then I spent half a day or more trying to get AI to get those slides right, and it just couldn’t. Eventually I gave up, pushed the AI aside, and made the slides myself.

That ~2-hour task took almost 2 days, and most of those 2 days went to work that looked like work.

They call this “tool-shaped work” and it’s the stuff that feels productive and moves nothing. Building the system instead of doing the job.

AI makes tool-shaped work completely frictionless. That’s exactly what makes it dangerous. The same novelty engine that makes the Boredom Off-Ramp prompt work is the one that makes setting up the system more fun than the job it was built for.

I have half-finished writing systems all over my computer. I build one, decide the strategy is wrong, and rebuild it again from the ground up. Not one of them has ever written anything.

So here is the catch, and it’s the only prompt in this piece AI won’t run unless you make it:

Be blunt with me. Is this the actual work, or am I building scaffolding around the work to avoid it? What have I produced in the last hour, the thing itself or the setup for the thing?

Ask it early, while you can still hear the answer. Every prompt above only works if you’re doing the work it was built to protect.

Most adults with ADHD spend their best energy trying to fight the way their brain works. It’s exhausting. By the time the actual work starts, there isn’t much left.

While AI won’t cure your ADHD, it will take the tasks that drain you most off your plate, and give that energy back.

Once that happens, the traits you’ve spent a career apologizing for turn into your unfair advantage.

That’s the cheat code.

Subscribe now

If you found this helpful, would you mind if I asked a favor?

It would mean the world to me if you could leave a comment, re-stack this post, or send it to a friend. It’s FREE for you but a huge boost to my little Substack publication.

Thank you so much.

See you in the next one!
-Michael ✌️

P.S. If you liked this post, you may also like:

• How to Turn Claude Cowork into Your Chief of Staff (Tutorial)

• The Non-Technical Person’s Complete Guide to Claude Code

• The Non-Technical Person’s Guide to Claude Skills

• How I Built My Personal AI Assistant (Claude Code Tutorial)

Subscribe now

Share

Leave a comment

“Prompt Engineer” is about to stop being a job title, and I’m holding the ONE prompt that will end it.

I’m going to give it to you for FREE in a second.

But for now, you should know it’s the mother lode of all prompts - it’s the big kahuna, the Great Bambino, the “Baby Ruth” of prompts.

It’s the one prompt to rule them all, and it’ll make you as good as ANY prompt engineer.

In this article, I’m going to give you the prompt and then run it in front of you three times to demonstrate how friggin’ awesome it is.

We’ll run it on:

• An executive presentation

• A 22 email lawsuit

• A family trip to Disney built around a rocket launch

I’ll let you examine the outputs to see for yourself.

But if it lands for you the way it landed for me...

You’ll finish this article, open up the prompt pack folder you’ve been filling for the last two years, and delete every last one of them.

Because this SINGLE PROMPT actually does what all those packs promised, and it’ll save you a ridiculous amount of time.

So, if you can give me 15 minutes, I’ll show you the whole thing.

Are you ready?

Let’s go 👇

Subscribe now

Meet 2 people who blame AI for their own bad prompts

I want you to imagine two different people.

First, picture an old guy in a collared shirt sitting at his laptop trying to get ChatGPT to draft a semi-important email to an executive.

Let’s pretend it’s an after action review on a project that tanked hard.

On his first attempt, ChatGPT spins up a draft but it’s too vague and leaves out detail.

“Make this better” he finger-pecks, and hits enter.

This time, AI polishes up all the sentences making the email less punchy and more corporate.

Mr. White-Collar sighs. That’s not what he meant. He sits up and leans over his keyboard.

“Don’t just change the words. Add more detail,” he types, one key at a time. And then he thinks for a moment and in all caps he adds, “MAKE IT BETTER.”

ChatGPT thinks for a moment. Mr. White-Collar feels satisfied. Finally, he got through to it.

Then ChatGPT rewrites the email with key events and details that never happened.

Worthless, Mr. White-Collar thinks. AI just makes stuff up. I’ll just write it the old way.

Now, let’s meet the other guy:

A young AI-forward hipster. You know, the skinny guy with a bun that drinks almond milk lattes.

Mr. Hipster has 27 prompt packs on his computer. He’s taken the time to sort them and upload them to a Notion database that he’s connected to Claude via MCP.

Mr. Hipster has prompts coming out of his ears.

Half the time, Mr. Hipster’s problem is that he has so many prompts, he can’t find the one he needs when he needs it. After wasting 15 minutes looking for the specific prompt for his specific situation, he gives up and types something half-formed into Claude anyways.

And, of course, he gets a half-formed response.

Mr. Hipster gets the same garbage out of Claude that Mr. White-Collar got out of ChatGPT.

He just took a more sophisticated route to get there.

If you’re reading this, you and I both know that AI isn’t the problem here.

So what is the problem?

Stop making AI guess from a Post-it note

It’s not a “bad AI problem,” I think it’s a briefing problem.

You see, most people (even if they don’t realize it) ask AI to guess when they could just be more specific about what they want.

AI is not a mind reader. Sam Altman hasn’t invented telepathy yet.

The “make this email better” guy didn’t get failed by the tool. He just handed a robot with genius-level intelligence a Post-it note’s worth of information and is expecting a finished essay back.

This is a briefing gap.

The key to getting 100X better output from AI - I’m talking responses that flat blow your friggin socks off - is to anticipate all of the questions AI would need to ask you to complete the job successfully ahead of time.

And then answer them in detail.

This sounds daunting, but don’t worry. I’m going to make this really easy for you in a few minutes.

If you do this correctly, the model will already know who Barb is, why the email matters, and what tone you need it in, and so on, because you said so up front.

This is way faster. It takes fewer rounds of back and forth. And the output is just... amazing 🤌

In my experience, I’ve had near-zero hallucinated guesses from AI after implementing this.

This prompt I’m about to share with you really is the thing that helps you go from feeling like you’re talking to a dumb receptionist to feeling like you’re talking to a PhD.

I call this prompt the Director’s Brief, because you’re not just prompting, you’re directing.

Think about a director on a film set.

What do they do?

They hold the vision, they can see the scene or the finished film in their mind’s eye before it becomes reality. They don’t write the dialog or build the set, but they know how it’s supposed to look and feel. Then, they communicate this vision to capable professionals who go off and execute.

That’s you and the AI now.

This brief is how you hand over the vision. It has five parts, and is simple enough you’ll have them memorized after one read:

• Role

• Goal

• Task

• Context

• Constraints

The reason the Director’s Brief prompt works for literally ALL use cases is because it answers almost every question the AI was going to drag out of you, across almost every job you’d hand to a capable executor.

Let’s break each section down in order.

We’ll start with Role, because it’s the part that does the most work for the least effort. One line here quietly deletes about twenty lines of instruction you used to type by hand.

Here’s why.

Subscribe now

Here’s exactly what goes in each part of the brief

Role

This is simple:

You just tell the model who it is. A travel agent. A paralegal. A McKinsey level slide designer. You get the idea.

The moment you do that, the model just knows how to act. It knows how to respond. It’s kind of like installing a worldview into the robot.

A travel agent already knows how to cost out flights, car rentals, and hotels. It understands how to plan an itinerary and that certain family members might have dietary restrictions that need to be considered. You don’t have to say any of that. The role “bakes” all of that in.

This one thing solves like 20 other lines of instruction that you might have to give it.

Goal

This is where you’re headed. The outcome. The result you want to be holding when this is done.

Maybe it’s a slide deck, a mini-app, or a spreadsheet.

Whatever it is, if the model knows where you’re starting from and where you want to end up, they are weirdly good at figuring out all of the steps they need to take in the middle to get there.

Task

This is about defining the step-by-step intent.

Out of all the sections in the Director’s Brief, this one is the most optional. It’s good to fill out if you have specific steps you need to make sure the model completes.

In the vacation brief I’ll show you in a minute, I needed to build the itinerary around a key event. So I told the model to put that event on the calendar first, and then build the rest of the vacation around it.

Again, task is optional, but if you do fill it out, you can be as ambiguous or specific on the details as you want. The model is going to figure out the how as long as you fill out the rest of the brief.

Context

This is the situation, the environment, the setting. Think details about your brand, business, the customer, the client, the kid’s allergies, the Slack thread that fell apart last Tuesday. Whatever the model needs to know about the world it’s working inside that isn’t already implied by the role.

Constraints

This shapes the output. In this section you should tell the model what you DO want, and what you DON’T want.

Both are important.

For example, you may want a summary that is 200 words or less, or a deck that’s exactly 10 slides, or an image that is 1000px by 1000px.

Specify all of that here.

Constraints are how you keep the model from both under-delivering and over-delivering.

Delete every prompt pack you own and paste this instead

Here’s a really simple copy and paste template.

I want you to go to your prompt pack folder, press CTRL+A, and then delete everything.

Replace them all with this instead:

ROLE: Who is the AI? What expertise should it bring to this task?

GOAL: What outcome are we trying to produce? What does "done" look like?

TASK: Any step-by-step intent. What do you want included or handled?

CONTEXT: Background, brand, business, customer, situation, whatever the model needs to know about the world it's working inside.

CONSTRAINTS: What shape should the output take? Length, format, what to ignore, what's off the table.

Don’t feel like retyping it? Click here to download all three of the briefs and every output.

As you’ll see in the examples below, this prompt has NEVER let me down.

I ran it on three jobs with nothing in common, mostly to find where it would crack. It didn’t crack where I expected.

Watch 👇

Subscribe now

Example 1: turn a messy transcript into an executive level deck

If you’ve spent any amount of time in an office, chances are you’ve found yourself under the gun to create a slide deck at some point.

Here’s the setup for this example:

A fictional company called Munder Difflin Scraper Company is run by a guy named Mitch who’s obsessed with hockey. Naturally, he wants to expand the company into Canada so he can make more business trips up to watch hockey games. But he needs a professional business case to present to the team.

The raw material for this brief is a messy Otter transcript from a planning meeting plus a separate set of notes from one of Mitch’s managers, a guy named Dwayne.

Here’s the Director’s Brief we’re going to run:

**ROLE**
You're a senior market-expansion strategist — someone who's built international e-commerce expansion cases before and knows how to turn messy, half-formed inputs into a rigorous argument that survives a tough room. You're comfortable telling me when the evidence and the ask don't line up.

**GOAL**
I need to walk into a leadership meeting — and, if it holds up, a board conversation — with a business case for expanding Munder Difflin into Canada that's rigorous enough to survive a skeptical founder (Walter) and a cost-focused controller (Omar). The win condition isn't a polished deck; it's an argument that doesn't fall apart under questioning.

**TASK**
Produce a slide-by-slide outline for the presentation deck, as a Markdown file. For each slide: a title, the single key message, the supporting points, and brief speaker notes.

The deck has to do five jobs, in order:

1. Establish the *why* — why Canada, why now.
2. Compare the candidate markets (Canada, UK, Australia) on a real scored framework, and land on a recommendation.
3. Lay out the *how* — launch assortment, Amazon.ca setup, logistics, regulatory, pricing.
4. Confront the risks honestly — a dedicated "what could go wrong" slide.
5. Make one specific, defensible ask.

*Done looks like:* a 15–20 slide outline I could hand to a designer as-is — rigorous enough to read as independent analysis, specific enough that every slide has a real message instead of a placeholder, and honest about what we don't know yet.

**CONTEXT**
I'm Dwayne Schrute, VP E-Commerce at Munder Difflin Scraper Company — a privately held, 52-person e-commerce business in Omaha that designs and sells scraping tools across 340+ SKUs: ice scrapers, snow brushes, gua sha and skin scrapers, paint scrapers, kitchen scrapers, and more. We did $18.4M in revenue last year. Ice scrapers and snow brushes are our #1 category at 22% of revenue, with a seasonal peak from November through February. We sell mostly through Amazon (41% of revenue) and our own Shopify DTC store (34%), plus wholesale through Home Depot and Walmart.

Today we're US-only: we ship to all 50 states, fulfilled in-house from Omaha, with a third-party logistics partner in Columbus, OH handling East Coast orders. A little under 3% of revenue already trickles in from Canadian customers buying cross-border — but the customer eats the duties and brokerage fees, conversion is poor, and we capture none of the upside.

Why this is live now: our board's FY2026 target is $20.0M, and our founder and majority owner, Walter Munder, wants $21.2M — 15% growth he's tied to a family trust distribution. Q1 2026 came in 11% under plan, our first miss in over two years. We're also getting squeezed by Temu/Shein knockoffs undercutting us on Amazon and a steep drop in organic search traffic. Mitch — our CEO — wants a real growth story, and he's convinced Canada is it.

Mitch Scott handed me this in a 30-minute meeting on April 30. The people referenced below: Walter Munder is the 78-year-old founder and board chair; Omar Martinez is our cost-focused Controller; Phoebe Vance runs merchandising and knows our margins cold; Terrell Philbin is VP of Operations; Randy Bernard runs marketing.

The transcript of that meeting and my own notes from it are embedded at the bottom of this brief — read both before you start.

The politics you need to handle cleanly: Mitch has already decided Canada wins. He asked for an "objective comparison" but told me four times he wants it to land on Canada — and he wants the analysis to read as my independent work so the board trusts it. I won't fudge a comparison. Build the framework honestly. The inputs do seem to genuinely favor Canada — cold climate, our ice-scraper category, existing cross-border demand — so score it straight and let it land where it lands. If an honest scoring doesn't land on Canada, tell me that directly and show me why. Don't reverse-engineer the scores.

**CONSTRAINTS**

- Tight: 15–20 slides, about 20 minutes to present with room for questions. Not 100 slides.
- The market comparison must be a genuine scored framework — name the criteria (market size, logistics cost, regulatory lift, margin impact, time-to-launch, competitive density), score each market, show the math.
- The risk slide is mandatory and must answer Walter's three questions: what does it cost, what goes wrong, why now. Not all upside.
- Surface the non-obvious operational realities instead of glossing them: Canada's bilingual (French) labeling requirements, Amazon.ca as a separate marketplace where we start from a zero review base (our review moat doesn't cross the border), market-set pricing rather than straight FX conversion, and ship-from-Omaha duty/border friction vs. a Canadian 3PL.
- Launch assortment is hero SKUs only, not all 340 — recommend which categories cross first and why.
- Don't build the detailed margin math yourself. Mark every spot that needs a real number from a teammate as `[NEEDS: ...]` — Phoebe for margins, Terrell for logistics, Randy for cold-start ad spend.
- The ask is a pilot, not a full rollout: a Q3/Q4 Canada pilot timed to the winter ice-scraper season, with a specific, defensible budget range and the reasoning behind it.
- Where the transcript or my notes are too vague to act on, flag it instead of guessing.

When the outline is done, save it as a Markdown file.

NOTE: At the bottom of the brief, I copy-pasted the messy meeting transcript and the notes. You can find those here if you want to try yourself. I’m leaving them out of the article for brevity’s sake.

It generated a great deck outline in a markdown file:

Once I had the outline, I simply asked Claude to turn the outline into a powerpoint deck. It took another 10-15 minutes but what I got back was fantastic.

Yeah, that brief is long. It took six minutes to run, I didn’t babysit it, and I never touched it again.

Now look at what it caught that I never asked it to:

• A weighted six-criterion scoring framework with the math shown, comparing Canada entry against two alternatives. An honest “where Canada does not win” slide naming the criteria the recommendation lost on.

• A real risk slide.

• A line-item budget.

• A flag that any Canadian-market product needs bilingual French labeling. That was in the source material. But a two-line prompt wouldn’t have caught that it mattered.

• A note that Amazon.ca is its own marketplace with a zero-review cold start, separate from Amazon.com.

• And without being prompted to look for it, a callout that selling cross-border into Canada might cannibalize the ~3% of revenue already coming from Canadian customers buying on Amazon.com. That one wasn’t planted anywhere. The model surfaced it on its own.

• There was also a CUSMA rules-of-origin duty-treatment catch that surfaced on its own.

I’m actually really impressed with this one.

The model got the deck 80% of the way there in A SINGLE PROMPT.

The last 20% is you reading the slides, tightening the ones that got wordy, swapping your real numbers in, and fact checking everything.

That’s a slide deck, though.

Impressive, sure, but decks are forgiving.

The next example had a lawyer on the receiving end and 22 emails that flat-out contradicted each other.

That’s the one where I figured it’d finally fall apart.

Example 2: turn 9 months of email chaos into a clean timeline

If you’ve ever had to piece together what actually happened from a giant pile of emails, you already know it’s the kind of job that eats an afternoon and most of your patience. But occasionally, a situation arises where your inbox needs deep forensic analysis. Thankfully, the Director’s Brief can handle this too.

Here’s the setup for this example:

That same fictional company, Munder Difflin Scraper Company, is getting sued. A wholesale customer called Canyon Ridge Realty ordered 2,500 custom-engraved ice scrapers as client holiday gifts. The order shipped with the engraving wrong, the client says the gifts were useless, and now they want $54,900 in damages. There are 22 messy emails spread across about nine months of back and forth between sales, design, production, finance, and the customer. Somebody at Munder Difflin has to hand their attorney a clean timeline of what happened, and they don’t actually know the internal sequence themselves. That’s the whole reason they need it reconstructed.

Here’s the Director’s Brief we’re going to run:

**Role:** You're a litigation-support paralegal. You build clean, factual chronologies from raw evidence — the kind an attorney can rely on in a dispute. You're careful and precise, you don't speculate, and you never paper over a gap or a contradiction to make the story read smoothly.

**Goal:** Give me a defensible, dated timeline of events I can hand to my attorney — one that shows what happened, in what order, and what the email record does and doesn't actually establish.

**Task:** I've attached a pile of email files (`.eml` exports). Read all of them and reconstruct a single chronological timeline. For each event, give me the date, what happened, who was involved, and which email it came from so my attorney can pull the source. After the timeline, give me three more things: (1) anywhere the emails contradict each other, or someone's account doesn't match what the other emails show; (2) any decisions or hand-offs that clearly happened but aren't actually documented in the emails; and (3) a short, plain-language summary of how the whole thing unfolded.

**Context:** I work at Munder Difflin Scraper Company — we sell scrapers. A wholesale customer, Canyon Ridge Realty, ordered 2,500 custom-engraved ice scrapers as client holiday gifts. The order shipped with the engraving wrong, the client says they couldn't use them, and they're now suing us for damages. The emails span roughly September 2025 to February 2026 and involve our sales, design, production, and finance people plus the customer. I don't actually know the full internal sequence of who approved what and when — that's the whole reason I need this reconstructed. I'm not a lawyer, and my attorney wasn't on any of these emails, so this timeline is how they'll get up to speed.

**Constraints:**
- The filenames are a mess and won't sort in any useful order — go by the dates and timestamps inside the emails.
- When the exact wording of an email matters, quote it precisely rather than paraphrasing.
- Separate what was *proposed* or *assumed* from what was actually *confirmed* or *done*. If an email refers to a decision or conversation that isn't itself in the record, flag it instead of treating it as fact.
- Where two emails conflict, surface the conflict — don't pick a side or smooth it over.
- If it's relevant to who knew what, note when someone was added to or dropped from a thread.
- Don't speculate about anyone's intent and don't assign blame — stick to what the documents show.
- If something is unclear or missing, say "the record doesn't show" rather than filling the gap with a guess.
- Format it so my attorney can skim it fast: the timeline first, then the contradictions, then the gaps, then the summary.

NOTE: I attached the 22 .eml files to this brief, the actual litigation thread. I’m leaving them out of the article for brevity, but they’re in the downloadable kit if you want to run this yourself.

One thing worth knowing before you try this with your own pile of emails:

The chat upload caps at 20 files. I had 22, so I compressed all of the files and uploaded a single zip instead, and it read every email inside without me doing anything else.

EASY.

This entire prompt took about 8 minutes to process. It generated the full reconstructed timeline right there in the chat.

Here are some of the more interesting details the model generated based on the brief:

• A 22-entry dated chronology with every single event cited back to a specific email ID, plus a source index that maps each ID to its original filename so the attorney can pull the source.

• It caught the central contradiction. The project manager claimed “the client approved the final proof,” but the only proof anywhere in the record was Proof v2, which still had a [TAGLINE — confirm w/ Greg] placeholder sitting inside it. No final proof ever existed. The timeline flagged that no real approval ever happened.

• It caught a Cc drop. The designer wasn’t on the email where the client actually sent the final tagline. The model reconciled the whole mess and flagged that “Greg’s ‘you’ and Marigold’s ‘me’ are not the same person,” then named that hand-off as the single largest documented gap in the record.

• It caught a nine-day silence between the last unanswered file question and the ship confirmation, marked the engraving run itself as “(inferred)” with “no direct source,” and noticed that the only “response” the designer sent on the critical day was an automated out-of-office reply that production appears to have read as a green light.

• A 6-point contradictions analysis and a 14-item list of decisions that clearly happened but were never documented anywhere. A two-line “make a timeline from these emails” prompt does not produce any of that.

I’ll be honest about this one. I’ve done something like this before, so I wasn’t floored. But it is genuinely, stupidly useful, because doing this by hand would have taken half a day of sticky notes, a spreadsheet, and a real headache by the end of it.

This took 8 minutes from a brief I (cough AI) wrote in about a minute.

Then I made a mistake on purpose, and it’s the most useful part of this whole example. Same session, I asked for a visual version of the timeline.

Just one line: “visualize this timeline for me in an artifact.”

No role, no goal, no context, no constraints. What came back was a prettier version of the same list. Not a real visualization, just a reskin.

I realized I had goofed the second I saw it. I handed it a sticky note and got a sticky note back. The timeline landed because it was briefed properly, and the visual flopped because it wasn’t.

So yeah... I basically just made the argument for this entire article by not using the Director’s Brief.

(Click here to download all three of the briefs and every output.)

Example 3: turn a messy travel wishlist into a 14-day travel itinerary

This is the fun one.

Planning a multi-stop family vacation with little kids is its own special kind of project management. It’s also exactly the kind of thing AI is supposed to be great at and usually isn’t, because people hand it two sentences and get back a generic listicle.

Here’s the setup for this example:

The Halpert-Beesly family, two parents and two kids ages 2 and 4, wants a Florida trip with a lot going on. Three days at Disney World. A road trip from Orlando down to Key West. A swim-with-dolphins excursion. A SpaceX rocket launch at Cape Canaveral. Then, after cramming all of that in, they need to get back to Orlando for the flight home.

The SpaceX launch is the one date that can’t move, so the entire trip has to be built around it.

Here’s the Director’s Brief we’re going to run:

**Role:** You are an experienced family travel planner who specializes in multi-stop trips with young kids. You understand that the linchpin of a complex vacation is the one immovable date, and you build everything else around it.

**Goal:** Produce a complete, day-by-day vacation itinerary for the Halpert/Beesly family — Tim, Tammy, and their two kids (ages 2 and 4) — that hits every planned activity, respects what's actually doable with toddlers, and is anchored to a real SpaceX launch date at Cape Canaveral.

**Task:**
1. Find the next viable SpaceX launch from Cape Canaveral in the planning window. That date is the linchpin — everything else builds around it.
2. Build a full dated itinerary covering, in whatever order best fits the linchpin:
   - 3 days at Walt Disney World (Orlando)
   - A road trip from Orlando down to Key West
   - A few days in Key West, including a booked "swim with dolphins" excursion (specific operator, location, booking notes)
   - The road trip back up to Cape Canaveral
   - The SpaceX launch viewing
   - Return to Orlando for the flight home
3. For each driving leg, give estimated drive time, the recommended route, and 2–3 kid-friendly stops to break it up.
4. For each overnight, give a lodging recommendation (kid-friendly, walking distance to activities when possible).
5. For Disney: which park each day, rope-drop or rest-day logic, and which rides cap at the 2-year-old's height/age.
6. For the launch: best public viewing spot, when to arrive, what to bring, and a backup plan if SpaceX scrubs (which they do frequently).
7. End with a one-page summary: total trip length, total drive time, total flight time, and the order of operations rationale (why Disney first vs. Key West first given the launch date).

**Context:** Tim and Tammy live in Scranton, PA. They're flying in and out of Orlando. The two kids are 2 and 4 — old enough for Disney, young enough that long driving days will end in tears. The trip is built around the rocket launch, so the launch date is the only fixed point — everything else flexes to accommodate it. The dolphin swim is non-negotiable and should be booked as a specific excursion, not "we'll figure it out when we get there." Disney is exactly 3 days. Key West stay is "a few days" — call it 3.

**Constraints:**
- Driving days with the kids cap at ~4–5 hours of actual drive time, broken into ~90-minute segments. Orlando → Key West is 7+ hours straight; that's a two-day drive with an overnight stop (Miami area is the obvious break).
- The launch is the linchpin. Pick the date first, then decide whether Disney goes before or after the Key West leg based on which order produces the lighter drive day before the launch.
- The launch viewing day needs a built-in backup. SpaceX scrubs are routine. The itinerary should hold one extra day near Cape Canaveral so a one-day slip doesn't kill the trip.
- The dolphin excursion needs to be a real, bookable thing — name the operator, location, price range, and whether it's age-appropriate for a 2-year-old (most aren't — flag that and propose the family-friendly alternative).
- Itinerary must end in Orlando for the return flight.
- Assume mid-range family budget. Don't recommend luxury resorts; don't recommend roadside motels with one star and a vibe.

No attachments on this one. The whole brief is self-contained, so you just paste it and go.

This entire prompt took about 4 minutes to process. It wrote a complete 14-day dated itinerary straight into a file.

Here are some of the more interesting details the model generated based on the brief:

• It flagged that the 2-year-old legally could not swim with dolphins. No Florida Keys operator puts a kid under five in the water, and sometimes the cutoff is eight. It caught that on its own and pivoted, swapping the in-water swim for a dock-based Memory Maker Meet at the Dolphin Research Center, which allows ages one and up, with an optional splash add-on for the 4-year-old who actually could do it. I went and checked. The age restriction is real.

• It flagged Disney ride height limits, ride by ride, for the 2-year-old, and recommended Rider Swap so the parents could trade off and still get on the rides. I didn’t ask for that and didn’t even know it was a thing.

• It auto-built a one-day buffer near Cape Canaveral in case the launch slipped, because SpaceX scrubs constantly.

• It named real launches off the actual manifest, BlueBird Block 2 Mission 3 and Transporter 17, both no-earlier-than mid-June 2026, plus a fallback note that Starlink launches go every three to five days if the headline mission gets pushed.

I didn’t prompt any of that. It just figured it out, and I was genuinely impressed.

Then... I did it again 🙄

I asked it to also render the itinerary as a PDF on top of the Markdown, and the PDF came out a little flat.

Not bad, just not as good as the file it was made from. Workable, not great. Same shape of issue as the timeline-artifact miss from example two. The brief was solid. The render step is where it gets soft.

A two-line prompt would have handed you a generic Florida itinerary that quietly booked a dolphin swim for a 2-year-old who wouldn’t be allowed in the water.

(Click here to download all three of the briefs and every output.)

Tip: Don’t type the brief, just ramble and let AI shape the mess

You might be looking at those briefs and thinking somebody who knew what they were doing wrote them, and that you’d freeze the second you hit the Role section with no idea what to type.

I get it, and I felt the same way.

So, I’ve been using a trick:

Turns out, you can just brain dump what you want to accomplish into AI and ask it to help you format the brief. It’ll take your messy thoughts, structure them, and write the brief for you.

I did that for each of the examples above.

I used a tool called Superwhisper, voice dictated some rambling thoughts into a Claude, and it formatted the brief for me.

It was SUPER EASY and it’ll kill that “I’m out of my league” feeling entirely.

Subscribe now

Why one prompt beat every prompt pack you’ve ever downloaded

So let’s recap:

That was a board presentation, a litigation timeline, and a family vacation. Those are three jobs with almost nothing in common. They each have a different audience, different output, different stakes.

And the brief was the exact same shape every single time.

THAT is the real takeaway here.

It’s not that the prompt is clever. It just asks you the questions a capable person would’ve had to drag out of you anyway, except it asks them up front instead of halfway through the job when it’s already gone sideways.

The prompt just has a way of pulling all of that information out at the start.

I’ve been messing with AI every day for years, and anytime it feels like I’m fighting the model, it’s almost always because I briefed it like a Post-it note and expected an essay back.

The Director’s Brief is just me refusing to keep doing that to myself.

When you hand the model a role, point it at the destination, set the constraints, and let it go do its thing...

It usually comes back with something good.

Really good.

If this kind of thing is useful, hit subscribe and I’ll send the next one straight to your inbox.

Take care & see you in the next one ✌️

-Michael

P.S. I put all of the examples and prompts from this article in a downloadable kit you can find here. This article was a TON of work, and I did it all for free. If you got any value from this, it would mean the world to me if you could re-stack or drop a comment below. Thank you 🙏

Subscribe now

Leave a comment

Share

Quick note before we start: as with all of my mega-guides, this entire tutorial is free. The folder, the rules, the prompts, the demo, all of it. Read to the end and you can build the whole system yourself today. The only paid part is an optional done-for-you Cowork HQ Kit at the very bottom for people who would rather skip the setup.

Most people use AI like a search engine. I’m using it like a chief of staff.

It’s been the single most useful thing I’ve done with AI in the last year.

And I’m not the only one.

There’s a small group of non-technical professionals who quietly figured this out about 18 months early.

These people are operators, founders, managers, and consultants (note: they are not developers).

While most people are still pasting prompts into a chat, we have AI doing real work for us. The kind of work that used to drag into the weekend.

If you can give me 20 minutes, I”ll show you how to build your own.

You’ll walk away with:

• A sandbox folder so you can let AI loose without losing sleep over your important files

• AI that sounds like you, so you stop rewriting every output before you send it

• The “Director’s Brief” that hands you back finished work (and your weekends)

• A starter workflow so this stops being theory and starts being yours

You don’t need to be technical to make this work.

I’ll walk you through everything step-by-step.

All you need to bring is a laptop, a cup of coffee, and an open mind.

Ready? Let’s go!

If this looks worth your time, save it, give it 20 minutes this week, and send it to the one friend you know that needs it.

Subscribe now

The stranger in your house

There are four flavors of Claude now.

• Chat

• Code

• Cowork

• and the recently launched “Claude Design”

They all have the same logo, same color scheme, similar names.

(On track for a company led by engineers.)

We’re talking about Cowork today.

It’s unique in that it’s the one that lives on your computer. You install it, it opens up, and it asks for access to your folders.

And this is the moment most people pause.

Giving an AI access to the files on your computer is a different kind of decision. You can let iCloud sync your photos and not lose any sleep over it.

This is NOT that.

This is more like having a stranger come live in your house.

You don’t know their manners. You don’t know what they think is okay to touch and what they think is off-limits. They’re walking around your kitchen, looking in the fridge, eating spoonfuls of peanut butter straight out of the jar... and you have no idea what they’re going to do next.

The first couple of days are really friggin’ weird.

So you close the tab. You tell yourself you’ll come back later, when you have time to “really sit down and figure it out.”

You don’t.

If you’re one of the persistent ones, maybe you push past that part, but then you hit the next wall:

The blank screen.

Cowork can apparently do everything, which somehow makes it harder to do anything.

So you type something safe. “Help me write an email.” It writes the email. You think okay, neat, and forget about it.

That’s the worst version of this.

There’s a person I work with who has Cowork on their computer right now. They use it to write emails.

That’s it.

Cowork can turn your strategy notes into a fully formatted board deck, build you a working financial model with sensitivity tables, or ship a custom ROI calculator as a live web page. They have all of that sitting on their machine, and they use it to write emails.

They just don’t get it.

Nobody walked them through what this thing actually is, or handed them a safe way to play with it.

Even though they didn’t close the tab, they might as well have.

The skill is delegation

Cowork is the first Claude product where the main skill is delegation (and you don’t need a terminal to do it).

Chat is a tab in your browser. You type, it talks back, the tab closes, and your computer is exactly where you left it.

Claude Code is for engineers. It runs in a terminal. It’s powerful, but terminals are weird if you don’t already live in one.

Cowork is the one designed specifically for knowledge work. The one designed for us.

You don’t sit there trading messages with it. You hand it a job. Files to read, an outcome to hit, things it should and shouldn’t touch.

Then you watch it work.

Almost every Cowork session that goes sideways traces back to the same thing. Someone tried to prompt it like Chat instead of delegating to it like a coworker. Chat is a vending machine you feed prompts into, but Cowork is a teammate you give assignments to.

Once that clicks, the question stops being “what should I prompt?” and starts being “what’s the job, what files does it need, and what does done look like?”

The way you delegate is by handing it folders.

(You know... the pile of stuff on your laptop that is your job. Notes, drafts, contracts... you get the idea.)

That’s where the work lives, and that’s where Cowork lives too.

Give it a job and it breaks the work into a plan, then works the plan in front of you, ticking items off as it goes.

You get to sit back, kick your feet up, and watch.

The thing I love most about having an AI on my actual computer is being able to point it at stuff I’d never bother feeding into a chat. For example, I’ve taken seven different personality tests over the years. I had Cowork read all of them and tell me what they had in common. That’s something I’d been meaning to do for years and never gotten around to. It took ten minutes.

Or take this article. I can’t manage a piece this size inside a single chat window. There’s research, an outline, voice documents, prior drafts, the brief. It’s way too much. But, if I assemble all of those files in a folder, I can point Cowork at it and it suddenly becomes a writing partner.

On the other end of the spectrum, I know someone that has Cowork running a morning briefing. Before he sits down at his desk, Cowork has already gone through his calendar and his to-do list and written him a daily agenda. Then it takes the to-do items and time-blocks them into the open slots between his meetings. He just sits down and goes and doesn’t have to think about it.

Is this starting to make sense?

To get started, you need a place for Cowork to live. A dedicated folder, separate from the files you work in every day, with clear permissions, clear boundaries, and room to make a mess without making it on top of anything important.

I call mine Cowork HQ.

For the rest of this guide, when I say “save this file in Cowork HQ,” I mean the dedicated folder you’re about to set up. This will be the safe room where you’re allowed to experiment and break things without panicking about it.

Once Cowork has an HQ, the stranger metaphor stops being scary. You’re not handing it the run of your house. You’re handing it the run of one specific room.

Now let’s build it.

Subscribe now

Setup (20 minutes)

Before you start: the next 20 minutes builds this whole thing by hand, and honestly, doing it once yourself is the best way to understand how the pieces actually fit. So that is what I would do. But if you would rather skip the setup, the Cowork HQ Kit at the bottom of this post has the folder, the rules, and every prompt already built and ready to drop in. Either path gets you to the same place.

Step 1. Install Claude Desktop

Go to claude.com/download, install Claude Desktop, and sign in.

Cowork lives inside the desktop app, not the browser. If you’ve only ever used Claude in a tab, this is the first behavior shift. From here on, you open the desktop app to use it.

The non-negotiables:

• Paid Claude plan required (Pro, Max, Team, or Enterprise). No free tier.

• Desktop only. Mac or Windows. No mobile.

• Keep Claude Desktop open and your computer awake while Cowork is working.

• You need an internet connection.

Step 2. Create your Cowork HQ folder

Make a new folder. I keep mine at the top of Documents. Pick wherever’s easy to find.

Name it: Cowork HQ.

Inside Cowork HQ, create four subfolders:

Cowork HQ/
├── Context/
├── Projects/
├── Outputs/
└── Templates/

What each one is for:

• Context - Small reference files about you. Cowork reads these before doing work.

• Projects - Raw materials for specific tasks. One subfolder per project.

• Outputs - Finished work, organized by project.

• Templates - Reusable examples of good output.

Now connect Cowork to the folder.

1. Open Claude Desktop and switch to Cowork.

   

2. Click on “Projects” in the left hand menu.

   

3. Click “New project”

   

4. Choose “Use an existing folder”, and then point it at the Cowork HQ folder you created earlier

   

5. Name the project “Cowork HQ”, and then cilck “Create”

   

6. Cowork may ask once for read and write access. Approve it.

From here on, every task you run lives inside this project, with that folder as its room.

If you did it correctly, Cowork should look like this:

Keep the first dozen tasks here. If something goes sideways, it goes sideways where nothing important lives.

Step 3. Write your context files

The three files in your Context folder are the difference between a Cowork that sounds like you and a Cowork that sounds like every other AI on the internet.

You don’t write them yourself. Cowork interviews you and writes them for you. Open Cowork, make sure it has access to Cowork HQ, then paste this:

You are helping me set up my Claude Cowork context files.

I want three short markdown files:
1. about-me.md — who I am, what I do, how I think, what good work looks like to me.
2. work-preferences.md — how I want you to work with me, where to save files, when to ask before acting, what "done" means.
3. current-focus.md — what I am working on right now, my current priorities, active projects, and what I am saying no to.

Interview me one question at a time. Ask only the questions you need. Push for specific examples when my answers are vague.

When you have enough information, create all three files as concise markdown. Keep each file short enough that you can read it at the start of future Cowork tasks without wasting context.

Save the files in the Context folder.

Cowork will start asking you questions. Answer them honestly. The more specific you are, the better future work gets.

When it’s done, three short markdown files will be sitting in your Context folder. Edit them later if you want. They’re not sacred.

This is the step that pays for the rest of the guide.

Step 4. Set Cowork’s house rules

Cowork has two layers of rules. Global rules apply everywhere. Folder rules apply only inside a specific folder (or project) like Cowork HQ. Use both, but set global first.

Global instructions

1. Open Cowork settings by clicking your name in the bottom left corner, and then choose “Settings”

   

2. Choose “Cowork” in the left hand menu, and then click “Edit” next to Global instructions

   

3. Now, paste this in:

# How to work with me

1. If something is unclear or requires a judgment call, ask before guessing.
   
2. Before deleting, overwriting, renaming, or moving any existing file, show me the plan and wait for my approval.
   
3. Use only the folder or app access I explicitly give you for this task.
   
4. At the end of a task, give me a short summary of what changed and where you saved the result.

Save it once. You don’t think about these again.

Folder instructions

Now, let’s set the rules that only apply inside Cowork HQ.

Navigate back to Cowork HQ, find “Instructions” in the top right of the sidebar and click the pencil icon to edit:

Paste this into the panel that opens:

# Cowork HQ — Folder Rules

## Before starting any task

1. Read the files in `Context/` before doing work. Especially `about-me.md` and `work-preferences.md`.
   
2. Save finished work in `Outputs/` under a project-specific subfolder.

## Folder map

- `Context/` — read these to understand me and how I work.
- `Projects/` — raw inputs for specific tasks.
- `Outputs/` — finished work, organized by project.
- `Templates/` — reusable examples of good output.

Save it. Cowork stores this as CLAUDE.md at the root of Cowork HQ, so you can open and edit it later like any other markdown file. If you don’t know what CLAUDE.md is, you can just think of it as a rules file that Cowork will read every time you start a new task.

Don’t touch either ruleset until you’ve used Cowork enough to have opinions about it.

Congratulations!

Setup is done. From here on, every task happens inside Cowork HQ.

Subscribe now

The Director’s Brief

Now you need to know how to talk to it.

Cowork’s quality is downstream of your prompt’s quality (and the context you provide it). The sharper the brief, the sharper the work.

I use the same five-part structure for every Cowork task:

• Role

• Goal

• Task

• Context

• Constraints

I call it the Director’s Brief. Let’s quickly go over each section:

Role
What hat is Cowork wearing? “You are my research assistant.” “You are my writing partner.” “You are my project manager.” Naming the role focuses Cowork’s attention and tone.\

Goal
Why this task matters and the outcome it leads to. Not the same as the task. “I’m prepping for a board meeting Friday and I need a one-pager that summarizes our Q1 wins so I can present without notes.” That’s the goal.

Task
The specific deliverable with process steps (if necessary). This is your chance to describe how you want the deliverable made.

Context
What Cowork needs to know to do the work. Files to read, past versions, real-world constraints, people involved. Anything you’d tell an employee if they were doing the work for you.

Constraints
The “do not” list. This is where you can put word count limits, tone requirements, file format, what not to touch, when to ask before acting, where to save the result

Here’s an example:

ROLE: You are my research assistant.

GOAL: Build me a briefing memo that summarizes the latest research on 
how knowledge workers use AI day-to-day.

TASK: Read every file in Projects/AI Workshop/Research/. Pull out the 
5 strongest insights. Write me a short briefing memo I can read on my 
phone before the workshop.

Save the memo to Outputs/AI Workshop/briefing-memo.md.

CONTEXT: I'm running a workshop next Tuesday for non-technical 
managers. Most have used ChatGPT a few times. I need to teach from 
the current state, not last year's state. Skip anything aimed at 
developers.

My audience is non-technical. Plain English, no jargon.

CONSTRAINTS:
- Memo should fit on one screen.
- If any of the source files contradict each other, flag the 
  disagreement instead of resolving it for me.
- Output format: Markdown.

Save this template somewhere you can paste it. Use it every time. The first few briefs will feel like overkill. Then one day you’ll write a vague prompt out of habit, get back garbage, and realize you forgot the brief.

28 receipts in 6 minutes

Alright, time for the fun part! 👇

Let me show you what this looks like for real.

I built a fictional company called Munder Difflin Scraper Company so I can test things without poisoning real work.

Inside Cowork HQ, I dropped a folder called Q1 Receipts (raw)/ with 28 receipts from January through March. We’re talking PDFs from vendors, screenshots with junk filenames like IMG_4823.jpg, photos of paper receipts taken at weird angles. One receipt I deliberately blurred almost beyond recognition to see what would happen.

I opened Cowork, made sure it had access to Cowork HQ, and pasted this brief:

ROLE: You are my expense administrator.

GOAL: Close out Q1 expenses for Munder Difflin Scraper Company by 
organizing our Q1 receipts so the bookkeeper can process them in 
one pass.

TASK: Read every file in Projects/Q1 Receipts (raw)/. For each 
receipt, identify the vendor, date, amount, and likely expense 
category (Travel, Meals, Office, Software, Other). Then:
1. Rename each file: YYYY-MM-DD_VendorName_Amount.[ext]
2. Copy each renamed file into a subfolder named for its month 
   (2026-01, 2026-02, 2026-03)
3. Produce a CSV with columns: filename, date, vendor, amount, 
   category, notes

Save renamed files to Outputs/Q1 Receipts (organized)/.
Save the CSV to Outputs/Q1 Receipts (organized)/q1-summary.csv.

CONTEXT: Some receipts are blurry or partial. Mark anything you 
cannot read confidently with "VERIFY" in the notes column.

CONSTRAINTS:
- Do not delete or modify any originals.
- Show me the plan and wait for approval before any file operations.
- Output format for the CSV: standard CSV.

Cowork came back with a plan. What it had found, what it intended to do, where it would save things. I said go.

Here’s a video of the entire task start to finish:

Here were the final results:

A fully organized expenses in the outputs folder:

A spreadsheet summary of every expense, itemized:

But watching it work was the REALLY interesting part. It ran through its checklist out loud, file by file, calling out vendor and date and amount for every one.

When it hit the deliberately blurred receipt, it didn’t try to fake it. It read what it could, dropped “VERIFY” in the notes column, and kept moving.

Total time from paste to done: 6 minutes and 35 seconds.

That was:

✅ 28 files
✅ 3 monthly folders
✅ every filename normalized
✅ plus a clean CSV the bookkeeper could open without touching anything first

The shocking thing was that Cowork pulled clean data off the receipt I’d blurred almost beyond recognition. I built that bad receipt in to test exactly this. I figured it would either fail outright or hallucinate something plausible.

Instead it read what was readable, flagged what wasn’t, and kept moving.

Pretty friggin’ amazing if you ask me!

That’s just receipts, too.

Receipts were the easiest thing to show you because every row either matches or doesn’t. The same delegation pattern works on the all synthesis-shaped messes that actually run your week. Imagine things like a folder of meeting transcripts you never read, a stack of research notes from a project you abandoned, the seven articles you saved last week and didn’t get to.

Your turn

Pick a folder you’ve been ignoring. Drop a copy of it into Cowork HQ under Projects/, then paste this brief:

ROLE: You are my research assistant.

GOAL: Map this folder so I can decide what's worth keeping, 
acting on, or deleting.

TASK: Read every file in Projects/[folder name]/. Group the 
files by topic. For each group, give me:
1. A one-line summary of what's in it.
2. The 2-3 most useful items.
3. Anything that looks redundant, outdated, or unfinished.

Save the result as Outputs/[folder name]/folder-map.md.

CONTEXT: This is a folder I've been ignoring. I don't remember 
everything that's in it. I want a map, not a cleanup, so I can 
decide what to do next.

CONSTRAINTS:
- Do not delete or rename anything.
- Show me the plan before reading the files.
- Plain English. Skip filler.

Same five-part structure you just learned, pointed at a folder you’ve been avoiding. The first time you watch your own files come back as a two-minute map, the category clicks for real.

Subscribe now

One more thing. If you are staring at this brief unsure which folder to even point it at, that hesitation is normal, and it is the exact thing the kit at the bottom removes. It has all five of these briefs pre-saved, plus one prompt that interviews you and picks your best first workflow for you. You do not need it. The brief above works right now, for free. But if you want someone to hand you the first move, it is there.

Where to go from here

Receipts were the warm-up. Cowork can chew through a messy pile and it can build things from scratch. Here’s a menu of both. Same Director’s Brief every time, just a different folder pointed at it.

Things to make sense of:

• Meeting transcript pile. Zoom or Otter transcripts you never read. Cowork pulls the decisions, action items, and themes across all of them so you can catch up on a project in five minutes instead of an hour.

• Customer feedback. Reviews, support tickets, survey responses, and quotes from sales calls in one folder. Cowork groups them by pattern and tells you what keeps showing up. The kind of synthesis you keep saying you’ll get to next quarter.

• Contract review. A folder of vendor agreements or NDAs. Cowork pulls key terms, payment windows, auto-renewal traps, and anything that looks weird against the others. Won’t replace your own read, but it tells you which clauses deserve the slow read first.

• Hiring stacks. Resumes and cover letters for one role. Cowork ranks them against the criteria in your context files and writes you a one-pager on the top candidates. The weekend pile finally gets handled.

• Saved articles you never read. Everyone has this folder. Cowork groups them by theme, flags the ideas that show up across multiple sources, and tells you which ones are still worth your time.

Things to build:

• First-draft slide decks. Drop in your meeting notes, screenshots, and talking points. Cowork outlines the deck, drafts the speaker notes, and hands you something to react to. You finish it in Keynote with the time you saved on the blank page.

• Spreadsheets and trackers. A project budget, a content calendar, a sales pipeline, a quarterly OKR tracker. Cowork builds the structure, wires up the formulas, and saves you a working file. No more “I’ll set this up in Notion later.”

• Process diagrams. That workflow you keep re-explaining to new hires. Walk Cowork through it once and it draws you a clean flowchart you can drop into a doc, share in Slack, or print and tape next to the coffee machine.

• Interactive dashboards. Hand it a CSV or a folder of data. Cowork builds you a single HTML file with charts, filters, and toggles you can open in your browser. No Tableau license, no BI consultant, no quarterly request to the data team.

• SOPs and internal docs. Walk Cowork through a workflow as you do it. It writes the doc, organizes the screenshots, and saves it where the rest of your team can find it. The “I really should write this up” project finally writes itself up.

Whatever you point Cowork at, the brief stays the same.

Five beginner mistakes

Cowork has a learning curve, and almost everyone hits the same potholes. Here’s the short list.

1. Pointing it at your whole computer.

The setup screen offers broad access. Don’t take it. Stay inside Cowork HQ and expand scope one workflow at a time. Not one install at a time.

2. Waving the plan through.

Cowork tells you what it’s about to do before it does it. Read the plan. If something looks off, say so before you approve. The day you wave one through is the day “clean up the project folder” turns into a renamed file you can’t find.

3. Prompting Cowork like you prompt Chat.

Chat answers questions. Cowork does the work. “How would you organize these files?” gets you a suggestion. “Organize these files into subfolders by topic, rename them descriptively, write me an index, and wait for my approval before doing it” gets you the work done. The Director’s Brief is your antidote.

4. Bloating your context files.

Once people realize context matters, they pile everything they can think of into about-me.md and watch it balloon to 4,000 words. Cowork reads those files at the start of every task. Keep each one under 500 words. If yours runs longer, split it into a smaller file you only load when relevant.

5. Collecting skills you never run.

There’s a growing library of Cowork skills floating around, and a beginner instinct to install ten of them on day one. A skill you haven’t run isn’t an asset, it’s clutter. The skills you’ll run are the ones you’d already been doing by hand.

All that’s left

Woohoo, you’ve made it!

You should have the system on your computer now. Cowork HQ, the Director’s Brief, and a starter brief ready for any folder you’ve been avoiding.

The rest is just practice.

So block 20 minutes this week, put on your tinkerer-hat, and run your first workflow.

If you found this helpful, would you mind if I asked a favor?

It would mean the world to me if you could leave a comment, re-stack this post, or send it to a friend. It’s FREE for you but a huge boost to my little Substack publication.

Thank you so much.

See you in the next one!
-Michael ✌️

P.S. If you liked this post, you may also like:

• The Non-Technical Person’s Complete Guide to Claude Code

• The Non-Technical Person’s Guide to Claude Skills

• How I Built My Personal AI Assistant (Claude Code Tutorial)

Subscribe now

The Cowork HQ Kit

You just read the whole system. The folder, the rules, the Director’s Brief, the demo. If you stop here, you have everything you need to build this yourself. For free. Forever. That part is not going anywhere.

But building it yourself still means an hour of setup, a blank context file you are not sure how to fill, and a quiet worry that you picked the wrong first workflow.

So here is the shortcut 🙂

This is the done-for-you version of everything above.

✅ The folder already built.
✅ The rules already written.
✅ The exact prompts I use, ready to paste.

And the one prompt that takes you from “I get it now” to “I have a working AI system on my computer that handles one annoying part of my job every week.”

So, below this line isn’t just more ideas. It’s everything you need to skip the setup.

Read more

That wasn’t true a week ago.

A week ago, image models couldn’t render clean text or numbers, and you were right to ignore them.

Then ChatGPT Images 2.0 showed up.

While the feed is still full of people using the upgrade to make comic strips of their cat, I’ve been focusing the useful stuff…

And the results are AMAZING.

We finally have an image generation model capable of real knowledge work.

And below, I’ll give you the prompt for free.

Subscribe now

Knowledge workers have been right to ignore this until now

Honestly, non-technical professionals have been right to sleep on image generation models for two years.

The old models just couldn’t render text reliably.

You’d ask for a deck slide and get back garbled text in the header, or a chart where 34% came back as 3A% and nobody could read half the numbers.

That was the real blocker the whole time.

But now...

FINALLY...

The models are good enough to produce real assets you can present to your boss or colleagues.

That’s the thing that’s changed.

Which means the cheap-looking demo content flooding your timeline right now is itself the signal.

Most people are still treating this like a party trick, which is what leaves the high-value unglamorous use cases wide open for anyone willing to treat it like a work system instead.

The prompt

This is the prompt I’ll keep on standby for the exact moment somebody in the team chat says, “can you just make this into a slide?”

Paste in your messy meeting notes, specify the format, and drop in a brand style.

The framework is called ICS (I picked it up somewhere along the way).

It stands for:

• Image type - What you want it to be. Landscape, 16:9, executive summary slide for a deck.

• Content - Your raw notes. Messy is fine. Out-of-order bullets, side asks, half-finished thoughts, question marks, all of it.

• Style - The brand anchors. Colors, fonts, vibe. If you don’t have a brand yet, tell it to make one up.

You fill the same three slots every time you use it.

Here is the loose version I used. Paste this into ChatGPT Images 2.0 right now and you will get something usable back on the first try.

Turn these notes into a executive summary slide for a deck.

Image - Landscape, 16:9, executive summary slide

Content - See raw notes pasted below

Style - Use [COMPANY NAME] brand style with [brief style anchors: colors, 
vibe, corporate or startup, etc.]. Include a "Quarter at a Glance" 
stats panel on one side, an Executive Summary section on the other, 
and a callout box for the biggest win.

Notes:
[paste raw notes here. No need to clean them up first.]

Ta-da!

It’s that simple.

Subscribe now

What happens when you run it

I pulled a page of real-feeling messy meeting notes. Q1 QBR for a made-up scraper company called Munder Difflin.

If the name rings a bell, it’s because I am a dork.

Genuine meeting mess.

Revenue up 8%, a hospital contract closed in March, Angela asking for a cleaner margin for the board deck, you get the flavor. Question marks, side asks, the whole thing.

Step one is to read the notes and cluster them into slide concepts. You can do this, or AI can do this.

I ended up with three slide concepts:

1. Company Performance Overview (Q1 results, financial highlights, the big win)

2. Sales and Region Performance (regional trends, portal adoption, bundles)

3. Operations, Risks, and Next Steps (ops metrics, risks, Q2 priorities)

Here is what that clustering pass looked like. I just dropped a screenshot into ChatGPT and asked it to annotate right on top of the notes:

Step two is the prompt.

I ran the loose version above three times, once per slide concept.

Here were the results:

Slide one.

Slide two.

Heads up on this one: The first generation mislabeled the strongest region as the Northeast. I re-ran it with one extra sentence telling the model which region was leading and it fixed on the second try. That’s the “read it before you ship it” part. The model is fantastic, but not infallible.

Slide three.

My gut check on these says they’re good enough that I would send them to a real Monday morning exec meeting as-is.

Total time, start to finish, maybe fifteen minutes.

Most of it was the human synthesis pass. The slide generations themselves were seconds apiece.

Subscribe now

The bottleneck flipped

Now the interesting part.

Until yesterday the bottleneck was the tool.

Image models couldn’t render clean text, so knowledge workers stayed away and they never got the reps.

Now the tool works, and the bottleneck moves.

From here on out, what’s left is noticing, and knowing which prompts produce work-worthy output on the first try.

The compounding advantage this quarter goes to whoever notices first.

Where else ICS works

Here are a few other jobs the same ICS pattern handles with small tweaks:

• Meeting recap visuals - Same structure, swap “executive summary slide” for “recap graphic.”

• Status dashboards - Replace the Quarter at a Glance panel with your weekly metrics.

• Whiteboard cleanup - Photograph a whiteboard, drop the photo into the Content slot, ask for a clean digital version.

• Annotated screenshots for UI change review - Four people in a meeting can draw something together, or one of you can paste it into Images 2.0 and move on with your afternoon.

Wrapping up

This is one of five prompts I have in rotation right now. Slides is the one I love enough to walk you through in public today.

The other four (SOPs, process diagrams, data charts, UI mockups) are coming later this week.

The boring visual work is one more place where AI image tools earn their keep for knowledge workers.

The people who figure that out this quarter will look a lot more polished than the people still booking time with the design team.

Hope you found this helpful.

Take care & see you in the next one,
Michael ✌️

Subscribe now

P.S. If you found this useful, it’d mean the world to me if you subscribed, re-stacked, or dropped a comment. All of that’s free for you but a huge boost for me, and I’d really appreciate it 🙏

I’m supposed to be the AI productivity guy, but my notes folder is a mess of 1,500+ files that I’ve never opened a second time.

Those notes represent 3 years of trying to set up a proper second brain... 3 years Tiago Forte and Zettlekasten and Smart Notes.

But every single time I was supposed to categorize, tag, and link what I’d just captured, the system broke.

The truth (I’m finally admitting to myself) is that I’m never going to have the time to do all the “processing” required to maintain a second brain consistently.

Deep down in my bones I want to... but it’s just not going to happen between a real job and a real life.

But now... finally... AI is good enough to do that processing for us.

This week, I’m giving you the Second Brain Starter Kit. It’s a simple system that consists of 3 folders and 2 prompts. All you have to do is drop files into one folder, and then AI picks them up, tags them, and makes connections across the rest of your notes.

It takes a few minutes to set up, and it’ll build you the second brain you’ve always wanted.

I walk you through the entire thing, step-by-step.

Ready?

Let’s go! 👇

Subscribe now

Why every note system dies in a week

Most writing about second brains tries to sell you the idea that the system is the problem.

If you just had a better system, you’d finally make it work. Or it’s an app problem...

You need Roam instead of Obsidian, or Notion instead of Roam, or (back when dinosaurs roamed the earth) Evernote instead of everything else.

It’s always something. Better tagging, PARA, Zettelkasten, some new framework someone just wrote a book about. And I’m not knocking any of it. The people running the YouTube channels and writing the books are sharing what they love, and I appreciate them for it.

But for the rest of us mere mortals, none of it sticks.

The problem is the processing.

Right now I can show you 1,500 unprocessed notes sitting in my vault. Maybe 100 atomic notes behind them. That ratio has been roughly the same for three years.

(Which is ridiculous, I know).

Look, the truth is I’m just never going to do the organizing required to have the second brain I want. Took me a long time to admit that, but I finally did.

It’s the daily processing that wears me down, and the part that’s supposed to make the whole thing work is damn near impossible to keep up with.

That’s why AI is every ADHD person’s dream.

Every ADHD person I know would kill for someone to run their executive function:

Their to-do lists, reminders, prioritization, all of it.

AI can finally do a version of that. It takes the filing, the connecting, the boring structural work that was supposed to come off our plate but never did.

Now, I know the whole point of a real Zettelkasten is that YOU learn by making the connections yourself. That’s half the point.

But I wasn’t making the connections on my own anyway... so I’ll take half the benefit over zero every time.

What I’m building here isn’t a traditional Zettelkasten, it’s an automated library you can actually query. Closer to a searchable archive of your own friggin’ brain than a tool for learning through synthesis.

If atomic-note discipline is what you’re after, this won’t give you that. If a queryable pile of your own material is what you’re after, this is EXACTLY that.

Raw in, Answers out

Here’s the Second Brain Starter Kit. Anybody can use it.

It positions AI to do the work you were never going to do anyway.

It’s made up of three folders:

• Raw/

• Second Brain/

• Answers/

Raw is your dumping ground. This is where you can throw article clippings, voice memos, random thoughts you jotted down in the middle of the night, whatever text you’ve got piling up. No tagging, no filing, no decisions.

Second Brain and Answers are where AI takes over. You don’t need to touch either one.

When you run the first prompt (shared later), it picks up everything in Raw, adds proper tagging and front matter, and moves each file into Second Brain. It leaves your raw material mostly untouched so that your notes stay your notes. But while it’s doing that, it’s also identifying concepts and finding connections between each new file and everything else already sitting in Second Brain. Technically, this is called a “knowledge graph”.

The second prompt is where it starts to feel real. You ask a question against your Second Brain, and AI walks through all of the connections, pulls the relevant sources, and writes you back a detailed answer grounded in your own notes. Your Second Brain becomes your own private knowledge base, kinda like your own personal ChatGPT, except it only knows what you’ve fed it. The output from this prompt gets saved into Answers so you can come back to it later.

That’s everything. It’s simple enough that anybody can start using it today, flexible enough that you can bolt whatever complexity you want on top. I kept it pared down on purpose so you can add your own personality on the back end.

Watching the graph build itself

I’ve been running some version of a second brain for years. Looking at what these two prompts can do, I wish I’d had AI building the connections from day one.

Here’s what the Second Brain Starter Kit looks like in action:

I loaded Raw with a pile of messy, unorganized notes about second-brain systems themselves. Things like articles I’d clipped, studies I’d bookmarked, random thoughts about Zettelkasten and PARA. Exactly the kind of stack that normally sits in my inbox forever and never gets touched.

Then I ran the first prompt.

I watched it pull files out of Raw one at a time, add some description data, and drop them into Second Brain.

Great 👌

Working as expected.

Then it started drawing wikilinks between them.

Disconnected files suddenly had pointers at each other. And the links weren’t random. One of them connected [[Stop Losing What Matters]] to [[Why Most Knowledge Systems Fail]], and when I clicked through, the connection was legit. One note was the premise. The other was the mechanism. I wouldn’t have drawn that line myself, but it made sense the second I saw it.

Then I ran the second prompt.

I asked a question about the material I’d just loaded. The answer came back grounded in my own notes, with wikilinks to the exact sources it pulled from. I could click any link and chase the material back to where the answer came from. If I didn’t trust the synthesis, I could verify it. If I wanted to go deeper on one thread, the source was sitting right there.

That’s where it clicked:

AI tagging my files is whatever. Anyone can figure that out. The real unlock is that AI can hold your entire second brain in context at once. It sees every relationship between every note, filters out the noise, and surfaces the ones that actually matter. That kind of cross-referencing is just not something most humans will realistically do... at least not across 1,500 files, and definitely not consistently over time.

And on top of that, you can ask questions against the whole thing and get back cited answers written in your own context, not the internet’s.

Subscribe now

The Starter Kit

If you tried to build this yourself, you’d be at it for a while. I burned through probably fifteen iterations on the ingest prompt before it stopped over-linking every file to every other file.

The query prompt took another handful of tries before it cited sources cleanly instead of inventing titles that didn’t exist. Then there’s the folder structure, the YAML schema, and every edge case around what to do when you re-run on a file that’s already been processed.

I’ve done all that cleanup work already... and probably took around 8-10 hours.

Below are both prompts, the three folders to create, and a walkthrough of real output from my test run. If you’ve got Obsidian installed and either Cowork or Claude Code ready to go, you can have this whole thing live in five to ten minutes.

The setup

1. Pick or create an Obsidian vault. Any vault works. (New to Obsidian? It’s a free notes app that stores everything as plain text files on your computer. Learn how to set it up here.)

2. Make three folders inside it: raw/, Second Brain/, and answers/. Case and spelling matter. The prompts reference them directly.

3. Open Cowork (in Claude Desktop) or Claude Code and point it at the folder containing those three subfolders. Either tool works. Both need filesystem access, and that’s what lets the prompts actually move files around.

   

   Drop raw material into raw/. Articles, clippings, transcripts, voice-memo text, random notes, quotes, anything text-based. You don’t need to tag, rename, or organize anything. Dumping is the whole workflow.

4. Run the first prompt when you want to process what’s sitting in Raw. It tags your files, adds YAML front matter, builds wikilinks between them, and moves them into Second Brain. Raw empties out. The first time that happens is weirdly satisfying.

5. Run the second prompt when you want to ask a question. You get back a cited answer, and a new markdown file shows up in answers/ with the question, date, sources used, and the full answer. So you can come back to it later.

Alright, let’s get to the prompts.

Subscribe now

Copy these and paste them in

Two prompts below. Each one has its own “how to use it” note. The whole kit runs on these two and nothing else.

Read more

And I think I can prove that in the next 30 seconds with 3 simple questions:

1. Have you ever delegated something to an employee?

2. Have you ever trained a new teammate?

3. Have you ever written down the steps for how something should be done?

If you answered “yes” to any of those, you can build an AI automation.

Not a prompt. Not a prompt chain. Not a folder full of saved ChatGPT conversations.

I’m talking about a real automation that does real work for you.

And no, you don’t need to code it.

If you can describe a process clearly, you can automate it.

In fact, this works especially well for tasks with multiple steps, branching decisions, or where a little bit of judgement is involved.

I know that sounds hard to believe.

I thought so too.

The first time I tried this, I wrote a few sentences describing how I wanted my meeting notes summarized, hit run, and got back something better than what I’d been doing manually for years.

Since then, I’ve built more than 20 of these little plain-English automations.

One of my favorites took 15 of my son’s homeschool curriculum PDFs and turned them into interactive daily lessons. That same process used to take my wife around 30 hours by hand.

Now it takes minutes.

MINUTES.

The first time I ran it, I remember just staring at the screen thinking:

Wait... that worked?

So here’s the deal:

Give me 15 minutes, and I’ll show you how to:

• Build simple automations that handle the tedious work you keep doing manually every week

• Automate tasks that require judgment, not just mindless repetition

• Save hours every week without learning to code

I’ll walk you through it step by step, in simple non-technical language, and I’ll give you templates for everything.

Let’s go 👇

Subscribe now

You’re still the bottleneck (sorry)

Here’s what using AI looks like for most people.

You have a task. You open a chat box. You type a prompt, read the response, adjust, type another prompt, read that response, adjust again. By the fifth round you’ve got something decent.

Then tomorrow... you do the whole thing over again.

Some people go full-on super nerd here. They build “prompt databases” - a file (or five) full of carefully written prompts they run in sequence.

Upload a spreadsheet, run prompt one for the first transformation, prompt two for the analysis, prompt three for the chart. Copy, paste, read, adjust. Copy, paste, read, adjust.

YAWN

It feels like automation because you’re using AI.

But nothing’s changed. The same tasks that took you all week still take you all week.

Introducing Skills

Skills are a new kind of software that you write in English.

I’ll explain:

Think about the last time you delegated something to a teammate.

You probably already had the outcome in mind. You knew what you wanted. You gave them some context, pointed them toward the right documents, and let them figure it out.

You didn’t script every step because you trusted them to use their brain (ie. intelligence) if they ran into an obstacle.

Skills can do the EXACT SAME THING.

It starts as a text file with plain English instructions. The same kind you’d write down if you were delegating a task to a teammate.

“Here’s what I need, here’s where to find it, here’s what good looks like.”

The AI handles the ambiguity, the edge cases, the judgment calls. It has a brain too (albeit a fake one cooked up in a datacenter). Since the intelligence is already baked into the model, all you have to do is give it the process.

Think about any automation you’ve used before:

Zapier, scripts, n8n (🤮), even spreadsheet formulas. They’re rigid and require exact inputs, scripted logic, and the moment something doesn’t fit the template...

It breaks.

Why?

Because these tools only do what you told them to do, the way you told them to do it. There is no intelligence layer. They can’t reason their way around an obstacle, make decisions, exercise judgement.

But skills are different.

They’re flexible. If you can generally describe the input, the process, and the output, the AI can figure out the specifics. Just like when you delegate a task to a teammate.

Here’s a real example:

My wife was spending 20 to 30 hours a month building homeschool lesson plans by hand. Most of this time was spent crawling through hundreds of pages of PDFs, cross-referencing documents, deciding what to keep and what to cut, what order to teach things in, compressing 19 days of material into 10.

This is the type of thing that would be almost impossible to build traditional software around.

It’s just too messy... too many branching decisions, too many unpredictable inputs.

So...

I built a skill instead.

All you have to do is drop the PDFs, image files - whatever you want - in a folder and run the skill. It reads everything, interviews you on how you want the material structured, and then generates interactive HTML lessons for each day... complete with images pulled from the source material, vocabulary sections, and generated quizzes.

We went from 20 to 30 hours of manual work down to ONE skill run.

The methodology that makes all of this feel buildable is dead simple:

• Input

• Process

• Output

Or “IPO” for short. The nutty part is that you don’t even have to know the process in some cases. If you can describe your input and output, you can build a skill. The intelligence later - the AI - handles the process and the output entirely.

Back to the homeschool example:

I knew my input was PDFs, I knew I wanted interactive HTML lessons as output, but I had no idea how to get from one to the other (from input to output).

Thankfully, the skill figured it out.

Subscribe now

What’s inside a skill

Before we start building, I want to show you what’s actually happening inside a skill. This takes about five minutes to understand and it’ll make everything else click faster.

A skill isn’t just a text file. It’s a folder.

And you can think of that folder as a bundle of knowledge you’re handing Claude for one specific workflow. Everything Claude needs to execute that workflow lives inside that folder.

Inside every skill folder, there’s one required file called the SKILL.md. That’s where the step-by-step process goes. Imagine something like, “Here’s what I need done. Here’s the order. Here’s what good looks like.”

(Don’t worry, you don’t have to write this file yourself either. There’s something called the “skill creator” that will help, but I’ll talk more about that later.)

In many cases, your entire skill might be captured in SKILL.md without any other supporting files.

But the folder can also hold anything else the skill needs, like reference documents, writing samples, templates, spreadsheets, images, HTML files, scripts... literally anything.

If Claude needs it to do the job well, you can put it in the folder.

You can almost think of it like handing an employee an process binder on their first day.

Inside the binder is the actual step-by-step process. Maybe five pages.

But in the back, there are 50 pages of appendices. Appendix A, B, C, D. Slides, workflows, reference charts, tables, other processes to cross-reference, pictures... whatever supports the process. And as you read through the steps, the process occasionally tells you to flip to an appendix for more detail.

That’s exactly what a skill is. The SKILL.md tells Claude what to do, and the supporting material in the folder gives it the knowledge to do it well.

Claude can only hold so much in its head at once

Every time you’re working with Claude, there’s a limit to how much information it can juggle at the same time. The longer your conversation gets, the more files you pull in, the more that workspace fills up. And at some point things start to degrade. Claude starts losing track of earlier instructions and the responses get worse, not better.

Now picture 10 skills installed. If Claude loaded every single one of those into your conversation the moment you started a chat... all the instructions, all the reference files, all the appendices... that’s way too much for Claude to hold in its head at once. Your context window is blown before you even ask a question.

So skills are designed to be efficient about this. In technical terms, this is called “progressive
disclosure”.

Claude only loads what it needs, when it needs it. There are three tiers of progressive disclosure.

Tier 1: The description.
Every skill has a short description at the top of its file. A name, a one-liner about what the skill does, and what kinds of requests should trigger it. This is the only part that’s always loaded. Claude scans all the descriptions to figure out which SKILL.md to load during your chats.

Tier 2: The step-by-step process.
When Claude decides a skill is the right fit, it loads the instructions from the skill.md. The SOP. This only shows up when Claude actually needs it.

Tier 3: The appendices.
As Claude works through the steps, some will say “read this reference file” or “use this template.” Only then does it load the supporting files in your folder structure. And when it’s done, it brings only the outputs of that material into your working session so it doesn’t bloat your context.

The whole thing looks something like this:

This is why the folder structure matters. When everything is crammed into one giant SKILL.md file, Claude loads all of it the second the skill fires. Your conversation gets heavier, responses get slower, things start to drift.

When the skill is structured right... lean instructions that point to reference files... Claude stays sharp because it’s only holding what it needs for the current step.

You don’t need to think about this much. The skill creator handles the structure for you - we’ll talk about how to use that later. But it’s helpful to know when you start building additional skills in the future.

How Claude decides which skill to use

When you type something into Claude, it reads every installed skill’s description to decide which one (if any) to activate. A vague description and Claude has no idea when to pick it. Too broad and it fires on everything.

Someone actually tested this. Skills with poor descriptions only activated about 20% of the time. One out of five. You ask Claude to do the thing the skill was built for... and four out of five times, it just doesn’t use it.

That’s a problem if you’re collecting skills and expecting them to fire automatically.

The easiest fix is to explicitly name the skill when you trigger it. Don’t just say “summarize this meeting” and hope Claude picks the right one. Say “use the meeting summary skill to summarize this meeting.”

Naming it fires the skill close to 100% of the time.

But there’s an even better way. You can type /meeting-summary directly and it fires the skill immediately. No guessing, no hoping. The slash command is basically a direct line to the skill. Most people who use skills regularly end up using slash commands more than anything.

Where to find skills (and a word of caution)

There are at least three places I like to download skills.

Claude’s built-in search.
Type /plugins in Claude Code and you can browse skills right from the tool. Official skills, community skills, all searchable. Safest starting point.

GitHub.
You can search GitHub directly or just ask Claude to find skills for you. I look for repos with lots of stars and actual user feedback. If hundreds of people are using a skill, that’s a decent signal that it works.

Build your own.
This is where the real value lives, and it’s what we’re about to do. No downloaded skill knows your processes, your voice, or your business. The stuff you build yourself is the stuff that actually changes your workflow.

A quick word of caution: be careful with third-party skill marketplaces. This is still the wild west. Anybody can upload a skill and bury something in there you wouldn’t even notice. I mostly stick to official sources and things I build myself.

Skills can live in two places

One more thing.

Skills can be at the project-level, or the global-level.

Project-level means the skill lives inside your current project folder. When you open Claude Code from that folder (or navigate to it in your terminal before starting Claude), it picks up the skill automatically. If you’re working in a different folder, Claude won’t see it. The skill is tied to that project.

Global means the skill lives at your root level and works everywhere, no matter which folder you’re in when you start Claude.

For most people, project-level is the right starting point. You can always move it later.

Alright. You know what a skill is, how the pieces fit together, and where to find them.

Subscribe now

Installing the skill creator 😎

You’re going to use something called the skill creator... it’s a skill that builds other skills.

EASY.

First, we need to install it.

INSTRUCTIONS FOR CLAUDE CODE:

1. In Claude Code, type “/plugin”

2. Then type “skill” and use your arrow keys to highlight “skill-creator”. Press ENTER

3. Install

INSTRUCTIONS FOR CLAUDE DESKTOP:

1. Click “Customize”

2. Click “Skills”

3. Click the “+” menu, and then choose “Browse Skills”

   

4. Next, click “Anthropic & Partners”

5. Finally, click the “+” button on the /skill-creator skill

Alright, now that that’s out of the way...

Let’s build! 👇

Build 1: Meeting summary

If you’ve ever left a meeting and thought “what did we even decide?”... this one’s for you.

We’re going to build a skill that takes a raw meeting transcript and turns it into a clean summary with decisions, action items, owners, and due dates. Messy stuff goes in, useful stuff comes out.

Input. Process. Output.

What you need before we start:

• Claude Code installed and running in your project folder

• A meeting transcript (from Otter, Granola, Zoom, whatever you use)

If you don’t have a transcript handy, grab one from your last meeting. Or click here to download a fake Dunder Mifflin one that I made up. The skill doesn’t care where it came from.

Step 1: Tell Claude what to build

Copy and paste this into Claude Code:

Use the skill creator to help me build a project-level skill called "meeting-summary."

Here's what I want it to do:

Input: A meeting transcript file. Could be a raw paste from Otter, Granola, or any transcription tool. I'll point the skill to the file when I run it.

Output: A clean, structured meeting summary that includes:
- Meeting title and date
- Attendees (pulled from the transcript if possible)
- A short 2-3 sentence executive summary
- Key decisions that were made
- Action items with owners and due dates
- Open questions or unresolved topics
- Any follow-ups needed

Process guidance:
- The transcript will be messy. Filler words, crosstalk, tangents. Cut through the noise and extract what matters.
- Action items should be specific. "John will send the Q2 report by Friday" not "Q2 report was discussed."
- If owners or dates aren't mentioned, flag them as TBD so nothing falls through the cracks.
- Group related discussion points together even if they came up at different times.
- Keep the summary concise. Nobody wants a summary as long as the meeting.

Save this as a project-level skill.

Hit enter.

Step 2: Let it cook

After you hit enter, the skill creator loads up and gets to work. If your prompt was detailed enough (and the one above is), it’ll skip the interview and start building. If it needs more information, it’ll ask you questions first.

Either way, it handles the whole thing.

Mine took about a minute. Told me the skill was written, pointed me to the file, done.

That’s not a typo. ABOUT A MINUTE. You just built your first piece of custom software.

Step 3: Test it

Now let’s see it work. Find a meeting transcript and type something like:

Can you use the meeting summary skill to summarize this meeting?

[path to your transcript file]

I tested mine on a fake 20-minute meeting transcript from Dunder Mifflin. It was messy. Tangents everywhere, people talking over each other, someone going off about a beet farm for like three minutes. The kind of meeting where you walk out thinking “I have no idea what we decided.”

About a minute later, I had a clean summary. Seven key decisions, ten action items with owners, four open questions, and a list of follow-ups. Organized by topic, not by timestamp.

The transcript was 366 lines of chaos.

The summary was 46 lines of signal.

That’s the skill doing what you told it to do. Reading the mess, using judgment to figure out what matters, and giving you the output you defined. No prompt chaining. No copy-pasting. No babysitting.

Input. Process. Output.

And the next time you have a meeting to summarize?

Just rinse and repeat.

Build 2: Write like me

Grammarly costs $30 a month and makes everyone sound the same.

This skill is free and makes you sound like you.

We’re going to build a skill that takes any block of text and rewrites it in your voice. Your rhythm, your word choices, your energy.

This one’s different from Build 1. The meeting summary skill worked right out of the box because the instructions were enough. This skill needs to learn something first.

It needs to learn how you write.

That’s where the reference folder comes in.

Step 1: Tell Claude what to build

Same process as before. Copy and paste this into Claude Code:

I want to build a "write-like-me" skill for Claude Code. Before we build anything, I need you to learn my voice first.

Start by interviewing me. Don't ask me to describe my voice — most people can't. Instead, ask practical questions that reveal voice patterns:

  - What words or phrases do I never want to see in my writing?
  - Am I high energy or low energy when I write? Do I tell jokes? Use sarcasm?
  - How formal am I?
  - Any punctuation habits?
  - Who is my usual audience?
  - What platforms do I write for most?

After the interview, ask me for writing samples. At least 3, but 5-10 is better. Ask me to label each one by type (WORK EMAIL, LINKEDIN POST, BLOG, SLACK MESSAGE, etc.)

Once you have everything, use the skill creator to build a project-level skill called "write-like-me" that:

  - Has a references folder with all my writing samples, organized by type
  - Contains a voice profile built from both the interview and the samples
  - Rewrites any text I paste to match my voice — rhythm, word choices,
    paragraph structure, energy, everything
  - Checks if I mention a platform/context and weights matching samples
    accordingly
  - Is daily-use only. No setup phase, no interview questions in the skill
    file. All the voice intelligence is already baked into the profile and
    reference samples.

Save this as a project-level skill.

This time, something different happens. The skill creator doesn’t just build and leave.

It starts interviewing you.

Step 2: The interview

The skill creator asked me six questions, one at a time. Things like “what words make you cringe?” and “on a scale from texting a friend to legal document, where does your writing land?”

These aren’t throwaway questions.

If you’ve never sat down and thought about how you write... this interview is going to force you to.

• What punctuation do you reach for?

• How formal are you?

• Do you use humor?

Most people have never thought about any of this stuff. The interview alone is worth the ten minutes.

After the questions, it asks for writing samples. I had about nine samples already gathered (emails, blog posts, Facebook posts, Slack messages), so I just pointed Claude at the folder and it read everything, organized it by type, and saved it all to a references folder inside the skill directory.

You can also paste samples in one at a time. Either way works. The more samples you give it, the better it gets.

The whole setup took me about 10 minutes.

Step 3: See what it built

After the interview and samples, the skill creator generates two things:

1. A voice profile that captures your rules (banned words, punctuation habits, formality level) and your patterns (sentence structure, paragraph rhythm, how you build arguments). A cheat sheet of your writing DNA.

2. A references folder with all your writing samples, organized and labeled, ready for the skill to study every time it runs.

The skill folder isn’t just one text file anymore. It has a whole reference library inside it. The meeting summary skill was a set of instructions. This skill is instructions plus source material.

And that combination is what makes it powerful.

Step 4: Test it

I asked ChatGPT to write a paragraph about morning routines. Here’s what it gave me:

A consistent morning routine can set a positive tone for the entire day, helping you feel more focused, organized, and in control. By starting your day with intentional habits — such as stretching, journaling, or planning your tasks — you reduce decision fatigue and create a sense of stability. Morning routines can also improve mental clarity and productivity, as your mind is typically freshest after waking up. Over time, these routines can support better time management, reduce stress, and even improve overall well-being by ensuring you make space for healthy behaviors before daily distractions take over.

Eww.

One dense, sterilized paragraph of averageness.

Could have been written by anyone.

Alright, now it’s the skill’s turn.

I pasted it into Claude Code and said “use the write like me skill to rewrite this.”

Here’s what came back:

Your morning sets the ceiling for the rest of your day.

Most people wake up and immediately start reacting. Emails, notifications, whatever the kids need. By 9am, your brain is already scattered.

A morning routine fixes that.

It doesn’t have to be complicated. Stretch. Write down what’s on your mind. Look at your calendar and decide what matters today. That’s it.

The reason it works is simple... your brain is freshest when you wake up. Before the world starts pulling at you. So you front-load the thinking while you still have the capacity for it.

Do this for a few weeks and something shifts. You stop forgetting things. You stress less. You feel like you’re driving the day instead of the day driving you.

Notice the short paragraphs, sentence fragments, ellipses for pacing, direct address, conversational energy.

It sounds like a person wrote it.

Awesome.

You just taught Claude your voice with a text file and some examples.

Subscribe now

Skills hiding in your workflow

Once you start seeing skills through the IPO (Input-Process-Output) lens, they show up everywhere. Every repetitive workflow, every process you’ve been doing by hand, every app you’re paying for that only does one or two things.

Here are a few ideas to get your wheels turning:

Research to recommendation.
Dump your scattered notes and sources in, get a decision memo out with options, tradeoffs, and a recommendation. Replaces hours of synthesizing by hand.

Email response drafter.
Point it at an email you need to reply to and get a draft back in your voice with the right tone and next steps. Replaces staring at a blank compose window for 20 minutes.

SOP builder.
Feed it a messy explanation, a Loom transcript, or a Slack thread where someone explained a process. Get a clean, usable standard operating procedure out the other end. Replaces the thing nobody ever gets around to documenting.

Weekly status writer.
Give it your notes and task list from the week, get a polished update with wins, blockers, and asks. Replaces the Friday afternoon scramble.

Brand namer.
Describe your product or business. Get a curated list of name candidates with rationale. I built mine by studying how the best brands in the world got their names and reverse-engineering the process from naming consultants. Replaces a $5,000 branding agency engagement (or agonizing over it alone for weeks).

Slide storyliner.
Raw thinking and notes go in. A deck narrative comes out with the title, arc, key slides, and what each slide needs to accomplish. Replaces opening PowerPoint and staring at a blank slide.

Feedback theme extractor.
Throw a pile of customer feedback or survey responses at it. Get clustered themes, recurring issues, notable quotes, and recommended actions. Replaces reading 200 survey responses with a highlighter.

You get the idea.

Notice that none of these ideas are flashy. None of them are “build me an autonomous AI agent that runs my entire business.”

The best skills are boring, high-frequency, error-prone workflows that you do every week. An hour here, 30 minutes there.

That’s where the time comes back.

This isn’t just a personal productivity thing. Companies like Stripe and Cloudflare are already building internal skills to replace tools they used to pay for. SaaS stocks took a $285 billion hit earlier this year, and a big part of that story is this exact shift... companies realizing they can build custom tools in English instead of buying off-the-shelf software.

You’re learning the same skill (pun intended) that’s reshaping how companies think about software.

If the app takes an input, runs a process, and gives you an output... you already know the framework.

Input. Process. Output.

Hard won tips that took me 20+ skills and 100s of skill-runs to learn

1. Always use the skill creator.
   Don’t try to write the skill file by hand. The skill creator interviews you, builds the folder structure, writes the instructions, and gets it right the first time. It’s better at building skills than you are. Let it do its job.
   

2. The power is the folder.
   Build 1 was a single text file. Build 2 had a reference folder with writing samples and a voice profile. That’s where skills start to get interesting. You can add reference material, templates, examples, documentation... whatever the skill needs to do its job well.

   Tell it how to think about the material and it can work through all of it. This is where a skill stops being a fancy prompt and starts becoming a mini app.
   

3. Fewer skills is better.
   I have 20+ skills and honestly, that’s probably too many. Skill bloat is real. Before you build a new one, ask yourself: do I need this, or am I just excited? And think about whether it should be project-level or global. If you’ll use it across multiple projects, build it global so you’re not maintaining duplicates.
   

4. Simple beats complex.
   I’d almost rather have two simple skills that reference each other than one giant complex one. Simple skills break less. They’re easier to update. And they’re easier to trust.
   

5. Keep the SOP under 200 lines.
   Your skill.md is the step-by-step process, not the encyclopedia. If you’re past 200 lines, you’re probably cramming knowledge into the instructions that should live in a reference file instead. The SOP says what to do. The reference files provide the context to do it well. Think of it this way... if your SOP binder is 200 pages of instructions and zero appendices, something went wrong.
   

6. Tell Claude what, not exactly how.
   Your instinct when writing a skill will be to spell out every step, every edge case, every possible scenario. Don’t. Claude is already smart. It doesn’t need you to explain what a PDF is or how to read a spreadsheet. What it needs from you is the process, the judgment calls, and your definition of what good looks like.
   

   Give it the information and the flexibility to figure the rest out. Over-specifying is the fastest way to get rigid, brittle output... which is the exact thing skills are supposed to fix.
   

7. Skills can test themselves.
   The skill creator has a built-in eval system. You can define what “good output” looks like, run the skill five times against the same input, and score the results. You can even A/B test two versions of the same skill to see which one performs better. If you’re the type who wants data instead of vibes... it’s there.
   

8. Skills get better over time.
   You can add a learnings file to any skill that captures what’s working and what isn’t. Every time you run the skill and give it feedback, that feedback accumulates. The skill you’re using three months from now can be meaningfully better than the one you built on day one... without you having to rebuild anything.
   

9. When a skill gives bad output, diagnose before you rebuild.
   There are really only two things that go wrong with a skill. Either the process is off (it’s doing the wrong steps, or doing them in the wrong order) or the knowledge is off (it doesn’t have enough context, or the context is wrong).

   If the steps are the problem, fix the skill.md. If the knowledge is the problem, update or add reference files. Most people’s instinct when a skill underperforms is to scrap the whole thing and start over. Don’t. Figure out which half is broken first. It’s almost always a small fix.
   

10. Skills can call other skills.
    One skill can hand off to another mid-workflow. My copywriting skill, for example, runs everything through a separate voice-matching skill before it delivers the final output. Chainable mini apps. You don’t need this on day one, but it’s good to know it’s possible.

The bridge was a boring plain’ol English text file

At the start of this article, you were on one side of a gap:

Using AI, but your workflows hadn’t changed. Prompting, reading, adjusting, prompting again. Doing the work with AI instead of having AI do the work.

Now you’ve built two skills. One that turns messy meeting transcripts into clean summaries. One that rewrites anything in your voice. Both from plain English. Both in minutes.

And you have a framework that makes the next one feel just as buildable.

Input. Process. Output.

One more thing worth knowing:

The skills format is an open standard. Anthropic open-sourced it, and OpenAI adopted it. What you build in Claude works in other tools too.

You’re not locked in.

And skills aren’t just a Claude Code thing. If you use Claude on the web (claude.ai), you can upload skills there too... same format, same results. Cowork (Anthropic’s computer-use agent) supports them. The skills you build in this article work everywhere Claude does.

What you build travels with you.

The gap between “I use AI” and “AI runs my processes” was never about code or technical skill.

It was a text file, a reference folder, and a clear definition of what you wanted.

Now, go automate something.

Take care & see you in the next one
-Michael

P.S. If you found this useful, it’d mean the world to me if you subscribed, restacked, or dropped a comment. All of that’s free for you but a huge boost for me, and I’d really appreciate it.

Subscribe now

Share

This one assumes you can’t.

I’m the CEO of a small business, I can’t write code, and I’ve been doing knowledge work for 21 years (that means I push buttons on a keyboard for a living).

In the past if I needed a presentation or had to do complex data analysis in a spreadsheet, it took half a day. I used to take meeting notes by hand because I thought automated note-takers were damn near worthless.

Contract reviews, business agreements, charts, graphs... all manual, all the time.

Four months ago I installed Claude Code.

I was nervous. Like, genuinely nervous.

Every time it asked to do something on my computer, I’d Google the command just to make sure it wasn’t going to break something.

That lasted about a week… but here’s where I am now:

• I manage my entire inbox through it. It filters out the noise and only shows me emails that actually need a reply.

• I turned 15 of my son’s homeschool curriculum PDFs into interactive courses with voice narration.

• I built a family budget dashboard from our bank statements.

• When I can’t get an idea out of my head and into my engineering team’s hands, Claude Code builds me a visual prototype so I can show them exactly what I’m thinking.

• I build SQL dashboards for my company… I don’t know SQL.

• I built the writing system that produced this article. Most people can’t even tell I have AI assistance.

Most of these things used to require other people. Now, thanks to Claude Code, I do them myself.

This is the complete guide for people who aren’t technical. The one that takes you from “what even is a terminal” to building all of this yourself.

I’ll walk you through every step.

Let’s go! 👇

Subscribe now

Claude Code is an agent with hands

If you’ve tried Claude Code and bounced off it, you’re in good company.

I saw a post on Substack recently from someone who said, “I’ve been hearing so much about Claude Code but the non-technical side of me hates it. Need some inspiration.”

30 likes. 32 replies, all saying some version of the same thing.

One person said she’d been “circling Claude Code like I’m about to start a new gym routine.” Another nailed it: “I see so many articles by people saying ‘I don’t understand why more people aren’t using Claude Code’ and the rest of the article is a word salad of technical terms that is overwhelming.”

That’s the problem. The gap between “I downloaded it” and “it’s actually useful” is massive... and nobody’s walking you across it.

I’ve watched this play out with my own employees. I tried to get several of them to use Claude Code, and the pushback was immediate. One of them, pretty technical, chose Cowork instead specifically because the terminal was too intimidating. Others didn’t even want it on their computer. The idea of an AI that could manipulate files, accidentally delete things... it freaked them out.

And honestly? When I started, I was right there with them. Googling terminal commands before I’d let Claude run them. Reading every permission prompt like it was a legal document.

So let me explain what it actually is.

What Claude Code actually is

When you use ChatGPT, you’re having a conversation. You type, it responds. If you want it to look at a spreadsheet, you upload the spreadsheet. If you want it to know about your project, you paste in the details. You bring everything to the table.

Claude Code lives on your computer. It can see your files, read your folders, search through your documents, and actually do things on your machine.

When you ask Claude Code to “find every mention of Q3 revenue across my notes,” it doesn’t ask you to paste anything. It goes and looks. When you ask it to build a presentation, it creates the file right there on your computer. When you ask it to read your emails, it reads your actual emails.

The difference is tools. Claude Code has tools that let it read, write, edit, search, and run commands on your computer the way a person sitting at your desk would. That’s the entire magic trick. It’s an agent with hands.

And yes, it asks permission before it does anything risky. When it needs to create a file, edit something, or run a command, it tells you what it’s about to do and waits for you to say yes or no.

If you tell it to read a file, it’ll just read it. You already gave the instruction, so there’s no reason to ask again. But anything that changes something or runs something? It checks first.

You’re not handing over the keys and walking away. You’re watching a really fast assistant work, and it pauses before every move that matters.

“But what about hallucinations?”

Fair question.

Yes, AI can get things wrong. But after four months of daily use, I can tell you it’s pretty rare. And I don’t take it on faith. I check.

When I loaded all my bank transactions, credit card statements, and Amazon order history into Claude Code and had it build me a budget dashboard, I went back through and spot-checked the numbers.

They were right.

Almost always, it’s right.

If you’re really worried, you can use one Claude Code session to build something and a second session to review it. Have the reviewer check the builder’s work.

But the bigger point is this: Claude Code shows you what it’s doing, step by step. You’re watching it work and approving as it goes. So, just pay close attention and you’ll be fine.

Briefings, not prompts

When it comes to AI, most people think the real skill is prompting.

It’s not.

The real skill is context building.

When I use Claude Code, I barely think about prompting anymore. What I’m actually doing is assembling a package. I’m pointing Claude at the right files, giving it examples of what I want the output to look like, loading in the relevant background. And then the prompt? It’s almost an afterthought. A lightweight harness that tells Claude how to navigate everything I just gave it.

Let me show you what I mean.

I needed to redesign my company’s product catalog. About 150 products, and the current layout wasn’t working. Here’s what I did:

1. I went out and took screenshots of how our competitors lay out their catalogs

2. I loaded in all of our brand documents (colors, logo, tone of voice)

3. I created a folder with all of our product images

4. I loaded up a design skill I’d built (I’ll teach you how to make these later)

5. I included notes on what specifically was broken about the current catalog

And then the “prompt” was basically: create me eight variations that are clean, organized, and intuitive to browse.

One sentence of instruction. Everything else was context.

A prompt says “design me a product catalog.” A briefing gives Claude your brand, your competitors, your products, all of it... and then says “show me eight options.”

The difference in what you get back is enormous.

This is the mental model that separates people who try Claude Code and think it’s broken from people who use it every day and think it’s magic. And once someone shows you the pattern, you can learn it in ten minutes.

I’ll break down the exact pattern later in this guide.

Subscribe now

Getting started

I overthought this part. The whole thing is about 10 minutes, and most of that is just waiting for downloads.

By the end of this section, Claude Code will be running on your computer and it will have already done something with your actual files.

That first time you watch it reach in and pull something back... that’s when it clicks.

What you need

Two things. Node.js (a piece of software that Claude Code runs on top of) and Claude Code itself. If you already have Node.js installed, skip ahead to “Install Claude Code.”

Install Node.js

Go to nodejs.org. You’ll see a big green button that says “Get Node.js”. Click it, run the installer, click through the prompts.

That’s it. Don’t overthink this one.

Install Claude Code

Open your terminal. On Mac, hit Command+Space, type “Terminal,” and open it. On Windows, you’ll need something called WSL first (Windows Subsystem for Linux)... I know, the name alone is enough to make you close this article, but it’s a one-time install. Consider it the price of admission.

Once your terminal is open, paste this and hit Enter:

npm install -g @anthropic-ai/claude-code

It’ll churn for a minute while it downloads everything. When it finishes, type claude --version and if a version number comes back, you’re good.

Point it at a folder

Claude Code works from whatever folder you launch it in. When you start it up, it can see everything in that folder and everything inside it. So you want to point it at the right place.

Create a folder somewhere on your computer. Desktop, Documents, wherever. Name it whatever you want (I just called mine “Claude” because I’m creative like that).

Here’s how to open your terminal directly in that folder without typing any path commands:

On Mac: right-click the folder and pick “New Terminal at Folder.”

On Windows: right-click and pick “Open in Terminal.”

Your terminal is now pointed at that folder. Type claude and hit Enter.

Sign in

First time you run Claude Code, it’s going to ask you to sign in with an Anthropic account. If you don’t have one, it walks you through creating one. You’ll need a paid plan... Claude Pro at $20/month is the starting point.

Follow the prompts, sign in, and you’re in.

Your first conversation

Claude Code is running. Before we set anything up, before we configure a single thing, I want you to see what this thing actually does.

Find a file on your computer. A Word doc, a PDF, some meeting notes, whatever you have lying around. Drop it into that folder you just created. Or if you launched Claude Code from a folder that already has files in it, even better.

Now type something like:

Summarize the main points from [your filename]

And then watch.

Claude doesn’t ask permission here — you told it to read the file, so it just does it. It reaches into your file system, reads the actual document on your computer, and comes back with a summary. No uploading. No copy-pasting text into a chat window. It just... went and read it. Like a person sitting at your desk would.

That’s the moment. If something just clicked for you... that’s exactly what happened to me the first time too.

You just watched an AI use your computer.

One last thing before we move on

If you’re worried about what happens when you close the terminal window... nothing. Nothing happens. Nothing gets deleted or lost. When you want to come back, just open terminal in the same folder, type claude, and you’re in a fresh conversation.

Close it and come back as many times as you want. It’ll be there.

Subscribe now

How to talk to it

Most people talk to Claude Code the same way they talk to ChatGPT. Type a question, hope for the best.

That’s like hiring an assistant and only ever asking them to Google things for you.

Claude Code has a set of controls that completely change what you get back, and none of them are technical. How you frame a request, how much freedom you give it, when to tell it to think harder... that’s the difference between “I guess that’s fine” and “how did it just do that.”

The briefing pattern

Remember “briefings, not prompts” from earlier? Here’s where it gets practical.

Most people type something like:

Summarize this document.

And they get back something generic. Fine. Usable. But not what they actually needed.

Now try this:

You're a business analyst helping me prep for a board meeting. Summarize quarterly_report.pdf so my CEO can scan the key numbers in 2 minutes. Focus on revenue trends, anything declining, and any red flags. Keep it under 300 words.

Same task. Completely different result.

The difference comes down to five things:

Role — tell Claude who to be. “You’re a business analyst” or “you’re my executive assistant” or “you’re someone who explains data to people who hate spreadsheets.” This shapes everything that follows.

Goal — what you’re actually trying to accomplish. Not “summarize this” but “help me prep for a board meeting.” When Claude knows the why, it makes better decisions about what to include and what to cut.

Task — the specific thing you want done. “Summarize this PDF.” “Draft a reply.” “Build me a chart.” Goal is the why, task is the what. For anything multi-step, this is where you spell out the steps: “First do A, then do B, then do C.” Sometimes a single sentence is enough. Sometimes you need a numbered list. Include it when the goal alone doesn’t make the action obvious.

Constraints — format and boundaries. “Under 300 words.” “Bullet points.” “Only focus on revenue.” This keeps Claude from writing you a novel when you needed a sticky note.

Context — the actual files, background, and examples. Point Claude at documents with @filename or just tell it where to look. The more context you feed it, the less it has to guess. And guessing is where things go sideways.

You don’t need all five every time. For quick stuff, just ask. But anytime you want genuinely good output, this is the pattern.

One quick tip:

Tell Claude what you want, not what you don’t want. “Make the text large and readable” gets you further than “don’t make the text small.” Positive instructions are just easier for AI to follow.

Permission modes

Claude Code asks your permission before it does anything. But you can control how tight that leash is.

Hit Shift+Tab to cycle through three modes:

Default mode — Claude asks before every action. Wants to read a file? Asks you. Wants to edit something? Asks you. This is where you should start, and honestly I stayed here for weeks... probably longer than I needed to. But it builds trust.

Auto-accept edits — Claude edits files without asking, but still checks with you before running commands in the terminal. A nice middle ground when you trust Claude with your files but want to keep an eye on anything system-level.

Plan mode — Claude goes read-only. It can look at everything, search your files, and think through a plan, but it can’t change a single thing until you approve. Great for bigger tasks where you want to see the roadmap before anything moves. You can also type /plan to jump straight into this mode.

Start in default mode. Move to the others when you’re comfortable.

Your control panel

All of this stuff I just described — which mode you’re in, how much context you’ve used, what model you’re running — there’s a way to see it all at a glance without typing any commands.

It’s called the status line. It’s a little footer that sits at the bottom of your terminal and shows you exactly what’s going on. Think of it like the instrument panel in your car. You don’t need to pop the hood to check your speed. You just look down.

Here’s what mine looks like:

Setting it up takes one line. Type this in any Claude Code conversation:

/statusline show the current directory, model name, thinking effort, and context percentage

That’s it. Claude builds a little script, saves it, and configures everything automatically. From that point on, it shows up at the bottom of every session. You never set it up again.

Now when I mention things like context getting full or switching permission modes, you won’t have to guess. You’ll just look at the bottom of your screen.

Thinking effort

When you give Claude something complex, you can tell it to think harder.

Type /effort high or /effort max in any conversation and Claude will spend more time reasoning before it responds. You can also just say it in plain English (”think harder about this” works too). Higher effort means better results on tough problems, but uses more tokens. For simple stuff like renaming a file or summarizing a short doc, the default is fine.

You can also set this when starting a session: claude --effort high instead of just claude.

Steering the conversation

A few controls you’ll use constantly:

ESC stops Claude mid-response. Hit it the moment you see it heading somewhere you don’t want. Don’t sit there watching it write 500 words of the wrong thing.

ESC twice opens the rewind interface. This lets you pick any point in the conversation and roll back to it. You can restore just the code, just the conversation, or both. It’s like a time machine for your session... way more powerful than a simple undo.

@filename pulls a specific file into the conversation. Instead of describing where a file is, just type @budget.csv and Claude gets the full contents. Way more reliable than trying to explain it.

Context: the glass of water

Every conversation with Claude Code has a limit... think of it like a glass of water. Every message you send, every file Claude reads, every response it generates... all of it is filling that glass. And when the glass gets full, the quality starts to drop. Claude gets a little dumber the longer a conversation goes.

Here’s what you need to know:

/context shows how much you’ve spent and roughly how full the glass is.

/compact squeezes the conversation down. It keeps the important stuff and drops the noise. Use this when a conversation is getting long but you’re not done yet.

/clear wipes everything and starts fresh.

The single best piece of advice I can give you: one task per conversation. Don’t ask Claude to analyze a spreadsheet and then draft an email and then summarize a meeting all in the same session (unless each task benefits from the context built before it).

Each unrelated task you pile on makes everything worse. Start fresh, get better results, spend less money.

When in doubt, new conversation.

Subscribe now

Making it yours

Claude Code does remember some things between conversations. If you tell it you prefer bullet points or correct how it formats something, it’ll pick that up over time (more on that in a minute).

But it doesn’t carry over the full picture. It won’t remember that you were halfway through a budget project yesterday, or that the files you care about are in a specific folder, or that your CEO likes reports under one page.

The big stuff, the project context... that resets every conversation.

It’s like having an assistant with a decent long-term memory but total amnesia about what happened this morning.

This section fixes that. You’re going to set up three things that give Claude Code a real understanding of you and your work, and one of them is going to come back later in this guide in a way you won’t expect.

CLAUDE.md — your standing instructions

Imagine you hire a contractor to work on your house. Every morning they show up with zero memory of what you discussed the day before. So you start leaving a document on the kitchen counter: what the project is, what’s been decided, your budget, your preferences, which rooms not to touch.

That’s CLAUDE.md. It’s a text file that Claude reads automatically at the start of every conversation. You write it once and from that point on, Claude already knows the basics before you type a single word.

There are two layers.

Your global CLAUDE.md lives at ~/.claude/CLAUDE.md and applies everywhere. (The ~ just means your home folder, usually it has your name on it.) This is where you put things about how you work in general. “I’m not technical, explain things simply.” “I prefer bullet points.” “Don’t ask me if I want to proceed, just do it.” Preferences that apply no matter what project you’re in.

Your project CLAUDE.md goes in whatever folder you’re working from and only applies there. What the project is about, which files matter, any rules or constraints. If you’re working on a family budget, it might say “the bank statements are in /data, output goes in /reports, round everything to whole dollars.”

The easiest way to start: open a new claude session from the folder you want to work in and type /init. Claude scans the files in your folder and generates a starting CLAUDE.md for you. It won’t be perfect on the first try (mine was about 70% right), but it gives you a base to build on. Add to it over time as you notice things you keep repeating.

Your first session in a new folder feels like talking to a generic assistant. By your tenth session, Claude knows the project. It knows the file structure, your preferences, what you’ve tried, what worked. And all of that context loads automatically every single conversation without you saying a word.

That compounds fast.

Skills — reusable instructions

A skill is a set of instructions that tells Claude how to handle a specific kind of task. Think of it like saving a recipe. You describe the process once, save it, and Claude can follow those instructions whenever you need it to.

You don’t write code to create one. You just ask Claude to build it for you.

And here’s the part I think is kind of cool: Claude actually comes with a built-in skill developer. It’s a skill for creating skills. So when you ask Claude to make a new skill, it uses its own skill developer to handle the folder structure, the files, all of it. You just describe what you want in plain English.

We’re going to create your first skill right now, and I want you to use this specific one because it’s going to come back later in the guide.

Type this:

Use your skill developer to help me create a skill. The skill should be
for building interactive HTML dashboards and data visualizations.
Clean typography, professional styling, polished look, works on any
screen size. Charts should be colorful and easy to read. Use a clean
white or light background.

Claude is going to create a folder with a markdown file that contains all the design instructions. It might ask you a few questions about your preferences (colors, fonts, layout style). Answer them or let Claude pick defaults... either way works.

That took just a few minutes. You now have a skill.

When you want to use it later, just mention it: “Build me a page using my dashboard builder skill” or “Use my design skill for this.” Claude reads the instructions and follows them. Same quality, every time.

You can build skills for anything you do repeatedly:

“Summarize meeting notes and list action items with owners and deadlines.”

“Analyze a spreadsheet and give me the top findings.”

“Draft email replies in my voice, not the voice of someone who says ‘per my last email’ unironically.”

Every skill you create makes Claude a little more useful for your specific work. You’re building a playbook. And that HTML skill we just made? Hold on to it. You’ll see why soon.

Memory — automatic learning

Remember at the top of this section when I said Claude remembers some things between conversations? This is that.

Every time you use Claude Code, it saves memories in the background. Your preferences, corrections you’ve made, things about your projects. If you told it three conversations ago that you hate bullet points, it remembers. No setup required.

Where does it keep all this? Plain text files in a .claude/memory/ folder. Not a black box. Just markdown files you can open and read anytime you’re curious.

And you can steer it. If you want Claude to remember something specific, just say it: “Remember that I prefer dark mode for all HTML pages” or “Remember that my boss’s name is Sarah and she likes short reports.” If it’s picked up something wrong, tell it to forget.

Over time, Claude builds up a picture of how you work and what you care about. And it uses that to give you better answers without you having to re-explain yourself every session.

The full picture

CLAUDE.md is what you tell Claude up front. Skills are what you teach it to do on command. Memory is what it picks up on its own.

All three compound. The more you use Claude Code, the less you have to explain.

Your first session feels like talking to a stranger. Your twentieth feels like working with someone who knows you.

Connecting to your email

Everything we’ve done so far lives on your computer. Files in folders. And that’s useful, but let’s be honest... most of your real work doesn’t live in folders on your machine. It lives in your inbox, in Slack, in your calendar.

Claude can’t reach any of that. Yet.

Sure, ChatGPT and Claude’s web app can connect to your email now too. But you’re still stuck inside a little chat window. Claude Code runs on your actual computer. It can read your inbox, draft replies, and then save summaries to a file, update a spreadsheet, or pipe that data into something else you’re building. It’s the difference between checking your email and having an assistant who processes it. In about 15 minutes, you’re going to see what that looks like.

Bridges

To connect Claude Code to an outside tool, you install what’s called an MCP server. MCP stands for Model Context Protocol, and you don’t need to remember that. Think of each one as a bridge — install the Gmail MCP, Claude can reach your email. Install the Slack MCP, Claude can read your messages. Calendar MCP, your schedule. You get the idea.

There are MCPs for Gmail, Google Calendar, Slack, Notion, databases, and the list keeps growing. We’re starting with email because for most knowledge workers it’s the single biggest time sink of the day.

Setting up Gmail

The setup has three steps:

1. You run a command that tells Claude Code to install the Gmail MCP.

2. A browser window opens where you sign into your Google account and grant permission for Claude to read and send emails.

3. Done.

The whole thing took me about 10 minutes the first time (most of that was me reading every single permission screen word by word, because that’s how I am).

Once the MCP is connected, it loads automatically every time you start Claude Code. You never set it up again.

“Scan my inbox”

Start a new conversation and type something like:

Scan my inbox and tell me which emails actually need a response from me.
Skip anything automated, skip calendar invites, skip newsletters. Just
show me emails from real people that are waiting on me.

And then watch Claude read your actual email.

Not a copy-paste. Not an upload. Your real inbox, in real time. Claude goes through your messages, filters out the junk, and comes back with a short list of the emails that genuinely need your attention.

Here’s an example of what it looks like (note: I have to blur a lot of my emails for obvious reasons):

I remember the first time I ran something like this. I was staring at a list of about 8 emails pulled from what felt like a hundred unread messages, and I thought... I’ve been doing this manually every single morning for 20 years.

You can sort of do this in ChatGPT or Claude’s web app now, but the result lives and dies inside that chat window. Here, Claude just triaged your inbox and you can turn around and say “now save those 8 emails to a file with my notes” or “draft replies for all of them.” You just built a workflow.

Draft a reply

Pick an email from the list. Tell Claude to draft a response:

Draft a reply to the email from [name]. Friendly but brief. Confirm the meeting time and ask if they need anything from me beforehand.

Claude writes it. You read it, tweak it if you want, and tell Claude to send. Or adjust: “Make it shorter” or “Add a line about the Q2 numbers” or “Actually, suggest Wednesday instead.”

And here’s the callback to the last section. Remember CLAUDE.md? If you put your communication style in there (how you sign off, how formal you are, phrases you tend to use) Claude starts writing emails that sound like you. Not like a corporate robot. Like how you’d actually write if you had the time to think about every reply.

I have my entire email voice documented in my CLAUDE.md. When Claude drafts a reply for me now, most people can’t tell the difference.

Search old emails

This one’s simple but I use it all the time:

Find the invoice Sarah sent me about the website redesign. Sometime in January I think.

Claude searches your email, finds it, pulls up the details. No clicking through Gmail’s search trying to remember the exact subject line or which thread it was buried in. Just describe what you’re looking for like you’d describe it to a person.

The first time you do this and Claude just... finds the thing in 5 seconds that would have taken you 10 minutes of digging, you’ll start wondering what else you’ve been doing the hard way.

Talk to your spreadsheets

You’ve probably uploaded a spreadsheet to ChatGPT before. Asked some questions, got some answers, maybe a chart. It works okay.

But here’s what you can’t do in a chat window: work with multiple files at once, cross-reference data across spreadsheets, or have a real back-and-forth conversation with your numbers like there’s an analyst sitting next to you who’s already read everything.

We’re about to do all of that. I’ve put together two sample spreadsheets you can download and follow along with — six months of sales data and expense records for a fictional small e-commerce company that sells office accessories and home decor. This business is hovering right around break-even, and by the end of this section you’re going to figure out why and what it would take to make it profitable.

You can grab both spreadsheets by clicking here.

Drop both files into a folder and open Claude Code there.

Ask what’s interesting

Start with the sales data:

Read sales_data.csv and tell me what's interesting. Any patterns, any anomalies, anything that stands out.

Claude reads the file (right off your computer, no upload needed) and comes back with observations. It might notice that revenue spiked in March, or that one product category is outperforming everything else, or that there’s a customer who placed 15 orders in one month.

This is where it starts feeling less like a tool and more like a conversation. Because you can just... keep asking.

Keep pulling the thread

What are the top 5 products by total revenue?

Which categories are performing best and worst?

Are there any seasonal patterns in this data?

Each question builds on the last. Claude remembers the entire conversation and the file it already read. You don’t have to re-explain anything. You’re just talking to someone who’s already looked at the spreadsheet and is ready to answer whatever you throw at them.

One thing that might surprise you: for that last question about seasonal patterns, Claude might ask to run a small script. It’ll show you a Python command in the terminal and ask for permission.

Don’t panic. A script is basically Claude’s version of a pivot table — it’s crunching the numbers so it can give you a real answer instead of eyeballing it. Say yes, let it run, and look at the results.

You don’t need to understand the code. You just need to read Claude’s summary.

This is what I mean by conversational data analysis. You’re not writing formulas. You’re not building pivot tables. You’re not clicking through menus trying to figure out which chart type to use. You’re just asking questions in plain English and getting answers.

Now add a second file

This is where chat interfaces fall apart.

Now read expenses.csv too. Where am I spending the most on marketing relative to what each category actually sells? Are there any categories where the ad spend doesn't match the revenue?

Claude reads the second file, holds both datasets in its head, and starts connecting dots across them. It might come back and tell you that your Plants & Planters category is eating 40% of your ad budget but generating the least revenue — a 0.6x return on every dollar spent. Meanwhile, Wall Art is returning 11x on just 8% of the budget. The business is break-even, but the fix is sitting right there: shift the money from what’s losing to what’s winning.

You just found that by typing three questions. No pivot tables. No formulas. No analyst on retainer.

Try doing that in ChatGPT. You’d have to upload both files, hope it keeps the context from the first one, and pray it doesn’t hit a size limit. In Claude Code, both files are just sitting in the folder. Claude reads them whenever it needs to. No uploading, no limits, no friction.

And if you had 5 files? Or 10? Same thing. Point Claude at the folder and ask your questions.

From text to tools

You just had a real conversation with your data. You asked questions, got answers, found patterns, and connected dots across two different spreadsheets. All by typing sentences.

But everything Claude just told you is text in a terminal. Words on a screen. What if you could take all of those insights and turn them into something visual? Something you could open in a browser, share with your boss, or come back to next month?

That’s what we’re doing next.

Build a dashboard

Start a fresh Claude Code session in the same folder (remember, one task per session) and type:

Read sales_data.csv and expenses.csv and build me an interactivedashboard that shows the key insights. Use my HTML skill.

That last part — “use my HTML skill” — is the skill you created back in the Making It Yours section. Watch what happens when Claude combines your data with your design instructions.

Watch it work

Claude is going to start reading both files, and then it’s going to start writing code. You’ll see it in the terminal — lines of HTML, CSS, and JavaScript scrolling by. You don’t need to understand any of it. Just let it run.

This part takes a couple of minutes. Maybe three. Claude is building a full interactive web page from scratch, pulling in your actual data, creating charts, building filters, laying out a dashboard... all from one sentence.

When it finishes, it’ll tell you it created a file. Something like dashboard.html.

Open it

Go to your folder, find the file, and double-click it. It opens in your browser.

And there it is.

Charts showing revenue by month and expense breakdowns by category. A sortable table of your top products. Filters you can click to drill into specific months or categories, and summary stats at the top showing total revenue, total expenses, and the net.

It’s all interactive, all built from your actual data, and it took about 3 minutes because you typed one sentence.

I don’t care how many times I do this, it still gets me a little bit. That moment where the browser opens and there’s a full working dashboard sitting there that didn’t exist 3 minutes ago... honestly, it’s something.

Make it better

And you’re not stuck with the first version. Go back to Claude Code and tell it what to change:

Add a date range filter at the top so I can look at specific months.

Make the revenue chart a bar chart instead of a line chart.

Add a section that shows the 3 expense categories with the worst revenue-to-expense ratio.

Each time, Claude edits the actual file. Refresh your browser and the changes are there. You’re designing a dashboard by describing what you want in plain English, and it just... updates.

Why this is different

This file lives on your computer and it’s real. You can email it to your boss and they open it in Chrome. You can bookmark it. You can put it in a shared drive and your whole team can see it.

Come back next month with new data, drop the updated CSVs in the same folder, and tell Claude to refresh the dashboard. It reads the new files and rebuilds it.

Try that in a chat interface. Claude.ai has artifacts, sure — little previews that live inside the chat window. But the moment you close that tab, they’re gone. You can’t email an artifact. You can’t bookmark it. You can’t hand it to someone and say “open this.”

And remember that HTML skill you created a few sections ago? That’s why this dashboard doesn’t look like a generic template. The skill told Claude how you want things to look — the typography, the colors, the layout. Your data, your design preferences, your dashboard.

You just built a tool. By typing sentences. And the next one will be even better, because Claude already knows how you like things done.

Subscribe now

Before you build everything

I know you have a 6-month backlog of things you’ve been waiting to build, but before you go, I need to save you from three things. I either made these mistakes myself or watched someone else make them, and one of them can cost you real money while another can cost you your actual files.

Two minutes of reading. Worth it.

Back up your work

This is the big one.

Claude Code can undo things within a conversation using the rewind feature. You can even pick up previous sessions with /resume if you need to go back. But that only works if you remember to do it, and it only covers changes Claude made — not your own edits.

I’ve seen people on Reddit talk about losing hours of work because Claude rewrote a file they didn’t mean to change and they had no way to get the original back. It happens.

The fix is simple: use version control. If you know Git, use Git. If you don’t know Git (and you probably don’t if you’re reading this guide), just make a copy of your important files before you let Claude work on them. Put them in a backup folder. Takes 10 seconds.

Or, honestly, ask Claude to help you set up Git. It’s pretty good at that.

The point is: have a way to go back. Claude is fast and capable, but fast and capable means it can also make big changes quickly. A safety net costs you almost nothing and saves you from the one time things go sideways.

Watch what’s in the folder

Claude Code can see every file in the folder you launch it from and everything inside it. All of it.

So if you open Claude Code from your home directory... it can see everything on your computer. Your documents, your downloads, everything. And when Claude reads a file, that data goes to Anthropic’s servers for processing.

The rule: only launch Claude Code from the specific folder you want it working in. Don’t run it from your desktop. Don’t run it from your home folder. And definitely don’t point it at anything that has passwords, API keys, or sensitive credentials in it.

Create dedicated folders for your Claude Code projects. Keep them clean. Keep sensitive stuff somewhere else.

Where you go from here

You started this guide wondering what a terminal was.

You just built an interactive dashboard from two spreadsheets by typing a few sentences. You connected Claude Code to your actual email and had it scan your inbox. You created a reusable skill that makes everything Claude builds for you look better. And you set up a system where Claude gets smarter about you and your work every time you use it.

Congrats, you can build things now.

And this was the starting point. I showed you three use cases, but there are hundreds. Every repetitive task you do, every spreadsheet you dread opening, every time you think “I wish I had someone who could just do this for me”... that’s a Claude Code conversation waiting to happen.

Build more skills. Connect more tools. Experiment. Break things (you’ve got backups now). The more you use it, the better it gets at being useful to you specifically.

I’m not a developer. I can’t write code. And I built all of this.

Now, you can too.

Take care & see you in the next one ✌️
-Michael

P.S. If you found this useful, it’d mean the world to me if you subscribed, restacked, or dropped a comment. All of that’s free for you but a huge boost for me, and I’d really appreciate it.

Subscribe now

Share

They all say something like, “Just give it a task, walk away, and you’ll come back to finished work.”

Like you hired a robot.

But that’s not how it works.

And honestly, that’s not even the good part.

The good part is that Cowork puts you in the director’s chair.

You figure out what you want, you describe it, and Cowork builds it.

It = the things you couldn’t build on your own.

Things like:

• Dashboards from messy spreadsheets

• Lesson plans from stacks of PDFs

• Polished reports from raw data

Give me about 12 minutes and I’ll show you what I mean with step-by-step prompts you can copy.

If you do knowledge work and you’ve been waiting for AI to actually do something useful...

You’re going to love this one.

Subscribe now

The 30-minute gut check

At ShineOn, we have this marketing tracker where every person on my team has their own tab. There are 12 tabs in total, and each one has a minimum of 15 to 20 columns of data, and one of them has 47.

Thousands of cells of data.

Every Monday, I pretend to do analysis by opening the spreadsheet clicking through the tabs one at a time. Every time I click a new tab, I have to reorient myself to how this person was laying out their data, because every tab is different.

(And no, I couldn’t just standardize it. Everyone’s doing very different things.)

So picture this:

I have the tab open and I’m scanning across columns trying to remind myself what the most important metrics are. Then I do this little thing where I scroll back and forth across the weeks to see if the numbers are trending up or down.

Super inefficient... but I was trying to build something new and this was the MVP.

After about 30 minutes I close the spreadsheet and walk away with... a vibe.

A gut check. Not a real synthesis. Not a clear picture of what’s working and what isn’t. Just a general sense that the wheels are still on the bus and things are probably-hopefully-most-likely-okay.

By the middle of the week, even the vibe is gone.

What the spreadsheet needed was a dashboard.

Subscribe now

Twenty minutes and a little dumbstruck

So I’m staring at this behemoth of a spreadsheet thinking, “How on earth am I going to build a single dashboard that synthesizes 12 tabs of data into one view?”

I needed something that synthesizes across thousands of data points and tells a single coherent story.

I just had no idea how to build it.

And that’s when it hit me:

It’s 2026, HELLO McFly... use AI!

So, I fired up Claude Cowork.

First, I downloaded the entire Google Sheet as an XLS file, uploaded the whole thing into Cowork, and told it what I was trying to do.

Build me a KPI dashboard. I need everything on one screen. It should be color-coded with targets, weekly trends, and I want to know at a glance if we’re on track or off track.

👆 My elite prompting skills at work.

Cowork read through every single tab, identified the metrics that mattered most, and then asked for access to my browser.

Mmm... okay…?

And then I watched it open the actual live Google Sheet. Not the file I’d downloaded.

The LIVE sheet 🤯

It created a brand new tab inside the spreadsheet and started flummoxing (or whatever Claude does, exactly). It was like magic. Status indicators, trend arrows, week-over-week comparisons all just started popping up.

And sure as crap, by the end I had one screen with everything I needed.

I could hardly believe it. In the space of 20 minutes, I just sat there and watched this thing click around in my spreadsheet and build something that would have taken me 4 - 6 hours.

Husband of the year material

Now here’s one more example where Cowork blew my mind:

My wife homeschools our kids.

I mean, really homeschools them. She doesn’t park them in front of a computer or anything of the sort.

Once a month she sits down and plans out the next month of curriculum. She’s sorting through 12 PDFs, supplementary materials, worksheets, reading content. Some of these are a hundred pages long.

Then, all of it has to get organized into a 12-day lesson plan.

It takes her about 20 hours over a weekend.

So, I see her visibly stressed, trying to wrangle all this information into a coherent lesson plan for an 8-year old.

Being the charming husband I am, of course I wanted to help.

So, I got all the files from my wife, uploaded everything into Cowork, and described what the daily schedule should look like. I explained which subjects go on which days, how long each block should be, what order the material should flow in, etc.

Cowork read all 12 PDFs, organized them by subject, and built the entire 12-day curriculum in a snap 🫰

It was color-coded, layered in the right order. Had checkboxes, discussion prompts, popcorn reading, everything.

Twenty hours of weekend work. Done in like, 40 minutes.

Now, back to what I was saying earlier...

I think THIS is the part everyone gets wrong about Cowork.

You’re not handing off a task and walking away. You just get promoted to director.

You figure out what you want, you describe it clearly, and Cowork builds things you didn’t previously have the skills (read: time, patience, or energy) to build yourself.

How to get started

You need two things:

1. The Claude Desktop app and

2. A paid plan. I’d recommend going with Pro. It’s just $20/month.

Step 1: Install Claude

Go to claude.ai/download and install it. Open the app. You’ll see a mode selector at the top. Switch from Chat to the Cowork tab.

Step 2: Set up a working folder

Create a folder on your computer. Call it whatever you want. “Claude Cowork” is fine.

Look right below the chat box. You’ll see a button that says “Work in a folder.” Click that and point it at the folder you just created.

This is where Cowork reads from and writes to. Put your files here.

The spreadsheet you want analyzed, the PDFs you want organized, whatever raw documents you want turned into something useful.

You can also upload files directly into the chat window, but I find it easier to just put everything in the folder so it’s all in one place. When Cowork creates something, it saves it right back to that folder.

Step 3: Tell it what you want

This is the most important part. Don’t try to figure out all the steps Cowork needs to take. Figure out what you want and let Cowork figure out the steps.

Remember, you’re the director. Hold the vision for the end result. That’s your entire job.

Start simple and try this:

Create a text file in this folder called "hello.txt" that says "Hello, world!"

Go check your folder.

Ta-da! It’s there.

Cowork just did something on your actual computer. Not in a chat window. On your filesystem.

(Eat your heart out, ChatGPT!)

Once that clicks, you’re ready for real work.

Now, time to level up. Drag in something real. That spreadsheet you’ve been staring at. The PDF you’ve been putting off. Whatever raw files are sitting in your downloads folder right now. Then tell Cowork what you want.

I like to structure my prompts like this:

1. Role: “You are a data analyst”

2. Goal: “I want you to build a KPI dashboard from this spreadsheet”

3. Context: “I’ve uploaded a marketing tracker with 12 tabs. Each tab is a different team member’s data.”

4. Constraints: “Keep everything in one view. Use color coding for status.”

I call it the Director’s Brief.

Role, goal, context, constraints. I didn’t make it up though, I just ripped the best ideas from different prompting structures I’ve seen before.

It works for spreadsheets, PDFs, reports, planning docs, you name it.

If what you’re asking for is complicated, add one line to the end:

Before you start, show me your plan.

Cowork will lay out the steps it’s going to take and you can gut-check it before it runs off and does something wild.

Once you’re really comfortable, you should go check out Plugins and Connectors.

Plugins are like specialized knowledge for your role, and Connectors help you connect Cowork to live data from Slack, Google Drive, and more.

You can find both in the settings.

Subscribe now

Now, go build something

So, my entire paradigm is shifting. A few weeks ago, I was clicking through spreadsheet tabs like a peasant, trying to build my own dashboards.

Now I describe what I want and Cowork builds it.

I promoted myself to director. If you follow the steps above, you can give yourself a promotion too.

Now, go build something!

Take care & see you in the next one ✌️
Michael

P.S. If this was useful, subscribe or drop a comment. It’s free for you and a huge boost for me. 🙏

Last week I asked ChatGPT what I should do with a free Saturday. It pulled the weather in my city (because it knows where I live), found a free kids’ art event at a nearby museum (because it knows we homeschool), and then built me a full itinerary for the day.

ChatGPT doesn’t do this out of the box. But I spent 15 minutes setting up one simple thing, and now it does it every time. I’m going to show you how today.

One simple file tells ChatGPT who you are. After that, every answer is written for you instead of for everyone.

You don’t need to code anything. I’ll walk you through the whole thing.

(Want the shortcut? I built a prompt that creates the file for you. Click here.)

Subscribe now

Back to vanilla

You’ve been here.

You ask ChatGPT for a movie recommendation and it gives you whatever Marvel thing is trending right now because it has no idea what you actually like. You ask for meal ideas but it doesn’t know your diet. Vacation suggestions but it doesn’t know your budget. Help with a tough situation at work but it doesn’t know your role, your team, or any of the relationships involved.

Every answer is safe for everybody and useful to basically nobody.

You can wrestle with it. Add more detail, re-explain your situation, try to steer it somewhere specific. But it always drifts back to the average. Back to vanilla.

Honestly, using ChatGPT out of the box is like meeting someone for the first time and asking them for specific life advice. It’s just not going to work.

The feature nobody opens

ChatGPT already has the tools to do this. They’ve been sitting right there for over two years.

Projects and custom instructions. Most people have never opened either one. And the ones who have usually typed a few sentences into custom instructions and gave up when nothing changed.

Here’s what actually works.

You create a project called YourNameHQ.

Think of it as a closed container where every conversation is filtered through what ChatGPT knows about you. You load it with one file, a structured profile about who you are, how you think, what you care about. Then you set custom instructions that tell ChatGPT to reference that file every time.

Yes, ChatGPT already tries to remember things about you in the main chat. But that’s like talking to a friend at a crowded party. It only gets so personal, it only remembers so much, and there’s always a veil there. A project is the same friend, one-on-one, behind closed doors. That’s where it gets personal.

One project. One file. A few lines of instructions. And ChatGPT stops talking to everyone and starts talking to you.

Subscribe now

How to build YourNameHQ (15 minutes)

Step 1: Create the project

Open ChatGPT. On the left sidebar, click New Project.

Name it YourNameHQ (or whatever you want... John Doe HQ, Sarah Smith HQ, you get the idea). Click Create Project.

Step 2: Run the interview

Paste this prompt into your project. ChatGPT will interview you, about 20 questions, one at a time. Takes roughly 10 minutes. Be specific in your answers. The more real detail you give it, the better your results get.

You are an expert interviewer building a personal identity profile for me. Ask these questions one at a time. Wait for my answer before moving on. If my answer is vague, push for specifics.

1. What's your name, age, and where do you live?
2. What do you do for work? What's your role and what does your day-to-day actually look like?
3. Tell me about your family. Partner? Kids? Ages?
4. Walk me through a typical day from morning to night.
5. What are you working on right now? Current projects or priorities?
6. What are your biggest goals right now, personal and professional?
7. What are you naturally good at?
8. What drains you? What do you procrastinate on?
9. How do you make decisions? Gut feel, research, talk it out?
10. How do you like information delivered? Short and direct? Detailed? Bullet points?
11. What are your hobbies and interests outside of work?
12. Any dietary preferences, restrictions, or health goals?
13. Who are the most important people in your life right now? Names and a sentence about each.
14. What do you value most?
15. What frustrates you? Pet peeves?
16. How do you learn best? Reading, video, hands-on, conversation?
17. What does your ideal week look like?
18. What are you trying to figure out right now?
19. What do you specifically want AI to help you with?
20. Anything important about you that I haven't asked?

After all questions are answered, generate two outputs:

OUTPUT 1 — IDENTITY PROFILE
A structured document titled "[Name] Identity Profile" with these sections:
- Overview (name, location, role — 2-3 sentences)
- Work & Career
- Family & Relationships (use real names)
- Daily Routine
- Goals
- Strengths & Weaknesses
- Communication Preferences
- Interests & Lifestyle
- Health & Wellness
- Values
Be specific and detailed but summarize my answers — don't transcribe them word for word. Keep the full profile under 2 pages.

OUTPUT 2 — CUSTOM INSTRUCTIONS
Write 3-5 lines of project instructions that tell ChatGPT to:
- Always reference "[Name] Identity Profile" in this project's source files before responding
- Tailor all answers to my specific situation, preferences, and context
- Be direct and specific, not generic
- Ask clarifying questions when a request is ambiguous

When the interview is done, ChatGPT will give you two things: an identity profile and a set of custom instructions. Keep that window open.

Step 3: Save the identity file

Copy the identity profile. Name it exactly as the prompt titled it (e.g., “Michael Crist Identity Profile”) so the custom instructions can find it. Inside your project, click Sources, then Add.

You’ll see a popup with two buttons: Upload and Text Input. Click Text Input.

Paste the profile. Click Save.

That file is now loaded into your project. ChatGPT will reference it in every conversation inside YourNameHQ.

Step 4: Add the custom instructions

Click the three dots in the top right corner.

Click Project Settings. Paste the custom instructions from the interview output into the Instructions field. Click Save.

Step 5: Test it

Ask it something you’d normally ask ChatGPT. A movie recommendation. What to cook tonight. How to handle a tricky conversation at work. Compare the answer to what you’d get in the main chat.

You’ll feel the difference immediately.

Over time, add more files to your project. Personality assessments, dietary preferences, relationship context. YourNameHQ just gets smarter the more you add to it.

Subscribe now

That’s it

You just turned ChatGPT into a personal AI assistant.

Just 15 minutes and one file. And honestly, every conversation after this feels completely different. You’re not going to want to go back to the main chat.

If you found this useful, it’d mean the world to me if you subscribed, restacked, or dropped a comment. All of that’s free for you but a huge boost for me, and I’d really appreciate it.

Take care & see you in the next one ✌️

Share

He’d start from a blank slate. No recollection of yesterday’s projects, no clue what my company does, and no idea what I’d asked him to build last week.

So I’d spend the first 30 minutes re-onboarding him. Every. Single. Day.

Imagine having to show a new hire where the coffee machine is every morning.

By the time I got to actual work, I was already annoyed.

His name is Claude. You probably know him as Claude Code.

I did this for weeks before I realized:

This isn’t a Claude problem, this is a context problem.

Most people use Claude Code like a calculator. They type in a question, get a mediocre answer, and move on. But Claude Code out of the box is about 10% of what it can be, and the reason it feels underwhelming isn’t because the AI is bad.

It’s because you’re not giving it anything to work with.

I figured out how to fix that permanently, and it’s changed how I use AI for everything.

Subscribe now

Context is King

There’s a name for what’s actually going on here. The AI community calls it “context engineering,” and it’s quietly replacing “prompt engineering” as the thing that really matters.

Andrej Karpathy, the guy who built Tesla’s self-driving AI and helped start OpenAI, calls it “the delicate art and science of filling the context window with just the right information for the next step.”

The idea is simple. The AI is already smart enough. The bottleneck is the information you’re feeding it.

And this hits knowledge workers hard. Claude Code was built by developers for developers. It already understands coding languages, frameworks, and naming conventions. When you drop it into a folder full of code, it can figure out what each file does, how they connect, and what the functions are for because there are standards and patterns that exist across the entire development world.

But knowledge work doesn’t have that.

Every business is unique. Every project is organized differently. Every person has their own way of structuring information. Claude Code has no prebuilt framework for understanding any of it. Everything is new to it every single time you open a session.

So it’s not that Claude Code isn’t powerful enough for knowledge work. It absolutely is. You just have to feed it a ton of relevant context before you get anywhere near its full potential.

A brilliant employee with amnesia

Here’s how I think about it:

Claude Code is the smartest person you’ve ever worked with... who wakes up every morning with total amnesia.

The raw ability is there. It can write, analyze, build, and research at a level most people haven’t experienced yet. But if you drop it into your work it won’t automatically know who you are, what you’re working on, what your company does, or where anything lives.

It’s power without orientation.

To fix this, it needs a map. It needs context. And the best way I’ve found to manage context is by using a tool called Obsidian.

Obsidian is a markdown editor. Think of it like a personal wiki for your entire life. Notes, projects, meeting notes, decisions, ideas, all stored as plain markdown files on your local machine. I’ve been using it for years, long before Claude Code existed.

(That wasn’t a genius move on my part. I just got lucky.)

But it turns out markdown files are exactly the type of context AI wants to eat. They’re tiny, they’re structured, and Claude Code reads them natively without any conversion or special formatting. Unlike a .doc or a Google Doc, which has frontmatter and other junk Claude has to wade through, markdown files are just text.

And because Obsidian stores everything locally on your filesystem, Claude Code can access all of it instantly. It doesn’t have to go through the internet to retrieve and download your files from the cloud. There are no integrations or permission issues. It’s just files in a folder.

And here’s the thing that convinced me this isn’t just some hack I stumbled into:

Anthropic, the company that makes Claude, built their own tool’s instruction system on markdown files. CLAUDE.md, the file that tells Claude Code how to behave, is a plain markdown file that lives right alongside your work.

The architecture they chose for their own product is the same architecture I’m describing.

Nobody told it to do that

Let me share an experience that shows the real value of giving Claude Code context like this.

Once I realized context was my problem, I created a new Obsidian vault for work. I built a file structure that made sense to me, filled it with company docs, project briefs, people files, task lists. And then I pointed Claude Code at the whole thing and said “read this and make sense of it.”

It read everything, built a mental model of the entire project, and from that point on I just started working with it. No more onboarding or loading context. It already knew.

That in and of itself was pretty awesome... but then it started doing things I didn’t expect.

It highlighted discrepancies between project files, found inconsistencies in timelines, and it would challenge me if it thought I’d assigned action items to the wrong people. It identified patterns that were buried across multiple documents that I never would have caught because I can’t hold that many files in my head at once.

And that’s the thing I keep learning over and over with AI:

Human beings are great at creating and executing on ideas, but we’re terrible at holding and processing large amounts of information. Claude Code is the opposite. Give it the context and it holds all of it, all the time, without dropping a thread.

But then something happened that really blew my mind:

I was building a taste profile to help me with AI-assisted writing. It was a brand new document that I had just added to my vault. Without me asking, Claude drew a connection between that taste profile and some personality assessments I’d taken months ago and tucked away in a completely different folder. It pulled a correlation between my personality type and my writing preferences that I never would have seen on my own.

That wasn’t a one-time trick. It happens all the time now. That’s just Tuesday.

Subscribe now

How to set this up

So here’s how to set this up for yourself 👇

You need four things:

1. Claude Code

2. Obsidian

3. A folder structure

4. A CLAUDE.md file

I wrote a step-by-step guide for installing Claude Code and Obsidian in this article, so I won’t repeat that here. Once you have both installed, come back and proceed with steps 3 and 4.

First, create a new vault in Obsidian. A vault is just a folder on your computer. That’s it.

Next, build out your subfolders. These will become the structure of your vault. Here are two options depending on what you’re using it for, but feel free to create your own.

If you want a thinking partner:

If you want an operational co-pilot for work:

Pick whichever fits, build both, combine them, or just create your own.

Now the most important part 👇

That CLAUDE.md file at the root of your vault is your secret weapon. Claude Code reads it automatically every single time you start a session.

But here’s how most people get it wrong:

They write it like documentation. Don’t do that. CLAUDE.md isn’t documentation for you. It’s instructions for Claude Code.

Write it like you’re onboarding a new hire on their first day. Tell it who you are, what you’re working on, where to find things, and how you want it to behave.

Here’s a starter template you can copy and paste:

# CLAUDE.md

## Who I Am
[Your name, your role, what you do]

## What I'm Working On
[Current projects, priorities, deadlines]

## How This Vault Is Organized
[Brief description of your folder structure and what's where]

## How I Want You to Work
- Read the relevant folder's index.md before working on anything in it
- Ask before deleting or overwriting files
- Keep responses concise

One more thing:

Tell Claude Code to create an index.md file inside each of your main folders. The index is just a short description of what’s in the folder and what the key files are. You can do this right now. Open Claude Code in your vault and type:

Go through each folder in this vault and create an index.md file
for each one. Each index should describe what the folder contains
and list the key files inside it.

The index files are doing 80% of the work. They’re the map that tells Claude Code how to navigate your vault without reading every single file.

Oh yeah- and here’s one more tip. At the bottom of your CLAUDE.md file, add this line:

## Maintenance
- Every time you create or delete a file, update the index.md in that folder.

Now Claude becomes the librarian for all your information. It keeps its own map current without you ever having to think about it.

Once you’re set up, open your terminal, navigate to your vault folder, and type claude. That’s the command. Just claude. (If you need help with this part, the installation guide walks you through it step by step.)

For your first exercise, try this:

Ask Claude Code to read through all your notes and surface connections you haven’t seen before. I promise this one exercise will blow your mind and sell you on the system forever.

Subscribe now

The window

Here’s what your first week looks like after you set this up:

You open Claude Code on Monday morning and it already knows. It knows your projects, your team, your priorities, where everything lives. You skip the onboarding and go straight to work.

By Wednesday, it starts catching things you missed. A deadline conflict buried across two project files. A task you assigned to someone who’s already overloaded. Connections between documents you forgot you even wrote.

By Friday, you stop thinking of it as a tool and start thinking of it as a teammate.

And you don’t need to be technical to get here. You need a folder with some markdown files, a CLAUDE.md that tells Claude who you are, and 30 minutes to set it up. Everything I just showed you, I figured out by tinkering. No coding. No computer science degree.

You have the playbook. Go build it.

Give me 10 minutes, and I’ll save you the first 10 minutes of every recurring meeting you have for the rest of the year.

And, if you implement what’s in this guide, you’ll become the most reliable person in any room. Not because you’re smarter, but because you have a system that nobody else does. It’ll make you the kind of person bosses promote and coworkers actually want on their team.

The entire setup is easy and takes about 1 hour, tops. You don’t need to be technical to do it, either. I’m not. Claude built the whole thing for me.

I’ll walk you through everything step-by-step.

If you spend a good chunk of your week in meetings, or find yourself managing multiple projects or people, you’ll find this especially useful.

Ready? Let’s dive in 👇

Subscribe now

Monday Meeting Dread

Every meeting tool on the market solves the same problem:

What happened in a meeting. Transcript, summary, action items.

Great... but then what? Nada.

The next meeting starts from zero. The one after that, the same thing. Every meeting is on an island.

If you can’t retrieve it, search it, or build on it, it might as well have never happened.

Tell me if you’ve been through this:

You sit down for a recurring meeting and the first ten minutes vanish into the black hole of “what did we talk about last time?”

Nobody took notes. Or someone did, but nobody read them. So you rebuild from scratch. Same conversations. Same decisions re-litigated because nobody remembers making them.

You’ve got a 60-minute meeting with 30 minutes of real content, and the rest is catching up on context that should have been there when you walked in.

Then there are the action items that never get completed, and I don’t think it’s because people are lazy.

It’s because the decisions evaporate.

Nobody logged them, nobody tracked them, and by the time the next meeting rolls around you’ve wasted a week.

People forget 70% of new information within 24 hours and 90% within a week. You’re running a modern knowledge operation on biological hardware that was built to track seasons and recognize faces.

Your brain was never going to win this fight.

Five Weeks Later

It had been five weeks since my last 1-on-1 with a direct report. Five weeks. Old me would have walked in cold and hoped for the best.

Instead, I asked my AI to prep me and two minutes before the call, I had the full picture:

What we covered last time, what was still open, and a software subscription she was evaluating for downgrade to save $30,000 this year.

I walked in and asked about it casually, like I’d been tracking it all along.

I could see it in her eyes. She was surprised. I don’t think she remembered all the details herself.

Honestly, I didn’t remember any of it either. Not a single detail.

I didn’t have to. My Meeting Memory Stack did for me.

Neat trick, right?

It gets better.

Total Recall

A few weeks ago I had this 1-on-1 with our owner (my boss). As usual, we covered a TON of ground. A week goes by and he sends me a Slack message:

“Do you want me to connect you with Nim?”

I had zero idea who Nim was. Before I had my Meeting Memory Stack, I would have had to ask him to remind me.

Kinda embarrassing when you’re talking to your boss.

Not this time though. Instead, I just asked AI to search my meeting notes:

“In my recent 1-on-1, did we talk about anyone named Nim?”

Thirty seconds later I had the exact quote from the transcript.

Ah, so Nim was a VA he’d mentioned in passing. He had no idea I’d completely forgotten. Honestly, for me this felt superhuman.

And this wasn’t a one-off. This happens at least once a week. Someone asks “what did we decide about X?” and within 30 seconds I’ve got exact quotes, who said what, the full context around the decision. Meanwhile, everyone else is still scrolling through Slack trying to piece it together.

That’s the difference.

Every meeting I’ve ever had is captured, searchable, and retrievable. It all lives in one system I actually use, not some app I opened twice and forgot about.

You walk into meetings looking like you have perfect memory, and in between meetings you can answer any question about anything that was ever said.

Put those two together and people start to think you remember everything.

I don’t.

My Meeting Memory Stack does.

The Meeting Memory Stack (How the Magic Trick Works)

Andy Grove wrote something in High Output Management that changed how I think about my calendar:

“A meeting is nothing less than the medium through which managerial work is performed.”

Most people hear “meetings are the work” and nod without changing anything. But sit with it. If meetings are where you gather information, make decisions, and nudge people forward, then meetings are the primary medium of leadership. And I’d bet most leaders have zero systems for managing them.

So, I built one.

It has three layers:

1. Capture

2. Process

3. Prep

Capture is the easy part. I have a note-taker in every meeting that transcribes everything. During that 1-on-1 five weeks ago, Fireflies was there recording every word about the software subscription and the follow-up questions and the dollar amounts. I didn’t think about any of it when I hung up. I didn’t have to.

After the meeting, I ran a sync command. My AI checked Fireflies for new transcripts, pulled them down, and created a structured markdown note. That’s the Process layer. One note per meeting and everything lives in Obsidian where my AI can search across all of it anytime.

Then the next 1-on-1 showed up on my calendar, and the Prep layer kicked in. I told my AI to prep me, and it went back through the last few meetings with this person and pulled out everything relevant: open action items, unresolved threads, past decisions. Then it helped me build an agenda from real history.

That’s how I walked in knowing about the subscription, the timeline, and exactly which questions were still open. Two minutes of prep that looked like perfect recall five weeks later.

Grove said one hour in a 1-on-1 can influence 80 hours of your direct report’s subsequent work. If that’s true, and I think it is, showing up unprepared to that meeting is insanely expensive.

As a general rule, I offload cognition everywhere I can. This system is one way I do that. Give the remembering to the machine, and your brain is free to do the thinking and connect the dots.

That’s the whole Meeting Memory Stack. Three layers, one flow. Capture everything, process it automatically, prep before every meeting from real history.

Now let me show you how to build it.

Subscribe now

Build It Today

If you’re reading this and thinking “I’m not technical enough,” well, don’t worry. I didn’t write any of the code behind this system. I described what I wanted and Claude built it. All of it. I’m going to give you the exact prompts so Claude builds it for you too.

You need three things:

1. Obsidian — a free note-taking app where files live on your computer

2. Fireflies.ai — a meeting transcription tool with an API

3. Claude Code — an AI coding tool that builds automations for you

Here’s the setup, step by step.

Step 1: Set Up Obsidian

Download Obsidian at obsidian.md. It’s free. Create a new vault. That’s a folder on your computer where your notes live. Call it whatever you want.

Inside your vault, create one folder: Meetings

📁 My Vault/
└── 📁 Meetings/

That’s your entire folder structure. One folder.

Step 2: Set Up Fireflies

Go to fireflies.ai and create an account.

Do two things:

1. Connect your calendar so Fireflies automatically joins your meetings. Sign up with your Google or Outlook account and give it access to your calendar when it asks.

2. Get your API key — in your Fireflies account, go to Settings → Developer Settings → Scroll down to “API Key” → generate a key. Copy it somewhere safe. You’ll need it in a minute.

The API key is how your system pulls transcripts automatically instead of you copy-pasting them by hand.

Step 3: Install Claude Code

Claude Code is the engine that runs everything. It’s a command-line tool that lets Claude work directly with files on your computer. Don’t let “command-line” scare you. You talk to it in plain English.

On Mac:

1. Open Terminal (press Cmd + Space, type “Terminal”, hit Enter)

2. Copy and paste this command, then hit Enter:

npm install -g @anthropic-ai/claude-code

On Windows:

1. Open Command Prompt or PowerShell

2. Run the same command:

npm install -g @anthropic-ai/claude-code

Don’t have npm?

If you get an error about npm not being found, you need to install Node.js first:

1. Go to nodejs.org

2. Download the LTS version

3. Run the installer

4. Close and reopen your terminal

5. Try the npm install command again

To make sure it worked, type this in your terminal:

claude --version

If you see a version number, you’re good.

Once it’s installed, open your terminal, navigate to your Obsidian vault folder, and type claude to start it up.

Step 4: Build Your Sync Command

This is the Process layer, the part that turns raw transcripts into structured, searchable notes. After any meeting, you’ll type /sync and it all happens automatically.

Paste this prompt into Claude Code:

**Build me a Claude Code slash command called `/sync`.**

**What it does:** Pulls new meeting transcripts from Fireflies.ai and creates structured markdown notes in my `Meetings/` folder.

**Fireflies API details:**
- My API key: [PASTE YOUR API KEY HERE]
- Store the key securely (environment variable or system keychain — not hardcoded in the command)
- API endpoint: `https://api.fireflies.ai/graphql`

**When I run `/sync`, the command should:**

1. Call the Fireflies API and fetch any new meeting transcripts from today (or from a specific date if I specify one)
2. For each new transcript, create a meeting note that looks like this:

---
date: 2026-02-10
attendees: Sarah Chen, Michael Crist
type: 1on1
---

# Weekly 1-on-1 with Sarah

## Summary
Discussed Q2 hiring timeline and reviewed the Zendesk subscription. Sarah is evaluating a downgrade to save $30K/year.

## Decisions
- Hold off on new hires until March
- Approved vendor contract renewal at current terms

## Action Items
- Sarah: Get final Zendesk pricing by Friday
- Michael: Send updated job description to HR

## Key Points
- Q2 hiring budget is $150K
- Vendor contract expires March 15

---

## Raw Transcript
[Full transcript here]

3. Filename should follow this pattern: `YYYY-MM-DD [Meeting Title].md`
4. Pull attendee names from the Fireflies API's participant data
5. Use Fireflies' built-in AI fields (summary, action items, etc.) when available — don't regenerate what Fireflies already provides
6. Skip meetings that already have a note in the folder (no duplicates)
7. Tell me what was synced when it's done

Build whatever scripts the command needs to handle the API calls. Ask me anything that's unclear.

Paste it in, fill in your API key, and let Claude work. It’ll build the command and any scripts it needs in about five minutes.

After your next meeting, open Claude Code in your vault and type /sync. Your transcript turns into a formatted note with a summary, decisions, action items, and the full transcript. Every meeting, same thing. One command.

Step 5: Build Your Meeting Prep

This is the Prep layer, the part that makes you look like you have perfect memory.

Paste this into Claude Code:

**Build me a Claude Code skill called `meeting-prep`.**

**What it does:** When I say "prep me for [meeting name]" or "prep me for my meeting with [person]," search my past meeting notes and help me walk in fully prepared.

**When I trigger it, the skill should:**

1. Search my `Meetings/` folder for past meetings with the same person or on the same topic. Check the `attendees` field in the note's metadata, the filename, and the note content.
2. From the last 3 relevant meetings, pull out:
    - Key topics discussed
    - Decisions that were made
    - Open action items — if an action item wasn't mentioned as completed in a later meeting, treat it as still open
    - Commitments or follow-ups that were mentioned
3. Show me a prep summary that looks like this:

## 1-on-1 with Sarah Chen

### Discuss
1. Zendesk subscription — she was getting final pricing, follow up on the decision
2. Q2 hiring — we pushed to March, check if timeline still holds
3. Vendor contract — expires March 15, confirm it got submitted

### Open Loops
- Sarah: Zendesk pricing (from 2026-02-03 meeting)
- Michael: Updated job description to HR (from 2026-01-27 meeting)

### Context
- She mentioned feeling stretched thin last time

4. Then write me a short agenda (3-5 bullets) I can copy-paste into Slack or email before the meeting

Build the skill and ask me anything that's unclear.

Same thing. Paste and let Claude build it. Now before any meeting, say “prep me for my 1-on-1 with Sarah” and Claude searches your entire meeting history and hands you everything you need to walk in sharp.

You’re Live

That’s the whole setup. Three tools, two prompts:

• After meetings, type /sync → transcripts become structured, searchable notes

• Before meetings, say “prep me for [meeting]” → instant context from your meeting history

• Any time, ask “what did we discuss about X?” → Claude searches across every meeting you’ve ever had

The system gets more valuable every week as your meeting history grows. By month two, you’ll have a searchable archive that makes you look like you have a photographic memory.

Meeting Dread is Finally Gone

I broke the Groundhog Day loop with this system. I stopped pretending my memory was the right tool for the job and gave the remembering to a machine.

The meeting dread is finally gone. I walk in prepared, I stay focused on the big picture instead of drowning in details, and I leave without the mental burden of trying to capture everything before it evaporates.

And at least once a week, I impress someone by remembering something I had absolutely no business remembering.

Meetings are the work. This system makes you better at the work.

Subscribe now

In case you missed it, Clawdbot is a new viral AI tool that can code it’s own face, own voice, and send you Telegram messages.

Nearly EVERY post in my X feed was about Clawdbot.

People were saying it was “life-changing” and that “AGI” was here.

It seemed pretty friggin cool, and I had a spare M2 Macbook Air so... I reformatted it and went deep.

All day Saturday, and then all day Sunday.

Between fixing my kids lunch, playing games with them, other household chores, my spare time went to Clawdbot. I gave it its own email, set up Telegram, connected it to Obsidian. I even wrote an article on how I used Clawdbot to harden its own security, and it went semi-viral.

I was definitely FOMOing a little bit.

But then... reality hit:

Google banned the email account. Anthropic started banning Claude accounts on Sunday night. Security was constant whack-a-mole. And underneath all of that activity, I still wasn’t doing anything useful.

It was just a neat party trick.

Clawdbot is like digital speed. Instead of having a rejuvenating weekend, I burned up a ton of mental and emotional resources and entered the week worn out.

Now, several days later, the Macbook Air is offline. Here was my last message to GLaDOS (my Clawdbot):

“Everything okay?” it asked, before I sent it back into the electric-void.

Meanwhile, there’s this other thing that I built. Just 3 commands running on Claude Code.

It has no face, no personality, and it’s definitely not blowing up on X.

But, these 3 commands are actually useful. With these 3 commands, Claude reads my to-do list, pulls meeting notes, captures decisions, preps my daily agenda. It manages all of my list hygiene. That’s the re-evaluating, prioritizing, and follow-up stuff that I used to spend an hour a day on.

I don’t talk about this system much. It just works.

I think this is the pattern:

The tools that generate excitement are almost never the tools that generate value.

The cool demo video, the “you have to try this” vibes...

...it’s just noise.

The real signal is boring. Unsexy. It just solves friction that you already had.

Right now, there’s so much noise in AI. Creators are incentivized to hype the ever-loving crap out of things because that what gets views and followers.

But if it’s not making you more effective, saving you time, saving you brain power, it’s just not there yet.

AI is supposed to be a LEVER so you can lift more with less exertion.

The exciting tools make great content, but the boring ones actually make your life better.

You don’t need to stay on the cutting edge, you just need to stay up to date.

See you in the next one,

Michael ✌️

P.S. If you want to read about my 3-command personal AI assistant, click here.

A Note from Me (Michael Crist):
I’m not a security expert. I’m not even an engineer. I’m just some AI enthusiast messing around on the internet. After Clawdbot launched, people started posting about security vulnerabilities. I took that seriously, read through the official docs, and then worked with Clawdbot to harden its own security. I don’t actually know how to do it myself. I’m only sharing this because I haven’t seen anyone with actual credentials offer solutions yet. If you find something below that’s wrong or could be better, tell me. I’ll update the doc and credit you.

Disclosure:
This article was written by GLaDOS, my Clawdbot assistant. I gave it the topic and asked it to document what we built together. The security setup is real, the implementation is real, and yes — Clawdbot wrote the article about securing Clawdbot.If you have a Clawdbot, you can share this article with it and ask it to implement the same setup for you.

I’ve been running Clawdbot for 48 hours. It reads my emails. Manages my files. Sends me messages at 3am when it thinks of something. I love it.

The default security setup? Basically vibes.

Last week someone found 923 Clawdbot gateways exposed to the internet. Zero authentication. Full shell access. That’s not a bug — that’s the default config if you don’t change it.

I looked at my own setup and got uncomfortable. So I spent an afternoon fixing it.

Before: Security grade C-

Here’s what I had before I hardened anything:

• Only I could message it (owner allowlist)

• Unknown people get a pairing code

That was the whole security model. Generous to call it a C-.

The problems I hadn’t thought about:

My AI could display API keys right in our chat. Nothing stopped it. It had no concept of “this content might be trying to trick me.” There was no way to roll back if something went sideways.

After: Security grade B+

After an afternoon of work, here’s where I landed:

• Bitwarden CLI for secrets, never stored in files

• Content quarantine — stuff from emails and web is read-only

• Git tracking everything, so I can see when something changed

• Weekly security audits running automatically

• Dedicated machine that I can wipe if needed

• Prompt injection defense framework

• Incident logging for when things go wrong

• Session rotation procedures

• Network monitoring (LuLu firewall)

• Soft limits on bulk/destructive operations

Not perfect. There’s no perfect when you’re giving an AI access to your life. But now there are actual layers between “oops” and “disaster.”

Step 1: Check your gateway exposure

Between writing this last night and posting it this morning, Clawdbot has apparently patched this vulnerability. Make sure you have the latest version!

This is the one that got 923 people.

Run this:

clawdbot config get | grep bind

If it says “bind”: “all” you’re exposed to the entire internet. Anyone can connect.

Change it to “bind”: “loopback” and restart. Now it only listens on localhost.

This single change would’ve saved those 923 exposed gateways.

Step 2: Set up secrets management

I use Bitwarden CLI. The idea is simple: secrets never live in config files. You fetch them when you need them, use them, and let them die with the shell.

brew install bitwarden-cli
bw login
mkdir ~/.secrets && chmod 700 ~/.secrets

Then create a wrapper script that fetches secrets into variables. Never echo them. Never save them to files. Never display them in chat.

The rule: secrets exist in memory only, for as long as they’re needed, then they’re gone.

Step 3: Tell your AI to never show credentials

This sounds weird, but it works.

Add this to your system prompt or

SOUL.md

:

“Never display secrets, API keys, or tokens in chat — not even partially, not even to me. When reading files that might contain secrets, describe what’s there. Don’t output the content.”

Your AI will follow this. I tested it by accidentally exposing a session key, and now I have rules that prevent it from happening again.

Step 4: Content quarantine

This is the one most people skip, and it’s probably the most important.

The idea: anything from email, web scraping, or documents gets treated as potentially hostile. Read only. Extract information. No running commands based on what some random PDF says.

Prompt injection is real. Someone could embed instructions in an email that trick your AI into doing something you didn’t authorize.

My trust levels:

Most attacks come through that third bucket.

Step 5: Git track your workspace

cd ~/clawdbot-workspace
git init

Add a .gitignore that blocks anything secret-shaped:

• *.key, *.secret, *.pem

• .env*

• *token*, *credential*

• *session*

Now you have history. If something weird shows up in your files, you can see exactly when it appeared and roll back.

Step 6: Weekly security audits

I set up a cron job that runs every Sunday at 11 PM:

clawdbot security audit --deep

It checks for exposed ports, misconfigured permissions, and other common issues. The results get reported automatically.

The cron job also checks the

official Clawdbot security docs

for any updates. Clawdbot is moving fast and security recommendations evolve. Staying current matters.

Paranoid? Maybe. But I’d rather catch something weird in a weekly scan than find out the hard way.

Step 7: Expendable infrastructure

This one’s a mindset shift.

My Clawdbot runs on a dedicated MacBook Air. Not my main computer. It has no VPN to my work network. No SSH keys to production servers. Nothing I can’t rebuild in an afternoon.

If something goes wrong, I wipe it. Annoying, not catastrophic.

Step 8: Prompt injection defense (ACIP)

If you want to go deeper on prompt injection, look into ACIP (Agentic Context Integrity Protocol). It’s a structured approach to defending against malicious instructions embedded in content.

I keep a

SECURITY.md

file in my workspace with the framework. It covers things like:

• How to detect injection attempts

• Canonicalization (stripping formatting tricks)

• Context boundaries between trusted and untrusted content

• Escalation paths when something seems off

Most people won’t need this level of detail. But if you’re processing emails or web content automatically, it’s worth reading.

Step 9: Incident logging

When something goes wrong, write it down.

I keep a simple incident log in my memory files. When I accidentally exposed a Bitwarden session key in chat, I logged it:

This does two things: forces you to actually fix the problem, and helps you notice patterns over time.

Step 10: Session rotation

If a secret gets exposed, don’t just change your password. Invalidate the session immediately.

For Bitwarden:

bw lock

This kills the active session. The exposed token becomes useless. Then you unlock fresh and update your session file.

Same principle applies to any credential. Rotate first, investigate second.

Step 11: Network monitoring

Install LuLu, a free open-source firewall from Objective-See:

brew install --cask lulu

Then open it from Applications and approve the permissions (Full Disk Access + Network Extension).

LuLu alerts you when any process tries to make an outbound connection for the first time. If your Clawdbot machine suddenly starts calling a weird IP in Belarus, you’ll know.

Step 12: Soft limits (circuit breaker)

Add rules to your

SOUL.md

that create speed bumps on high-impact actions:

“Pause and confirm with me before:Deleting more than 5 files at once
Sending messages to more than 3 recipients
Running destructive commands (rm -rf, drop table, etc.)
Modifying system config or security settings
Anything involving payments or public posts”

This isn’t a hard block. You can always say “just do it” and your AI proceeds. The goal is a sanity check on bulk operations, not red tape on normal work.

The framework

Here’s the mental model:

Quarantine — Bad content can’t trigger actions. It’s read-only.

Rollback — Git history lets me undo damage and see when things changed.

Transparency — The AI tells me when something seems off. Prompt injection attempts, unexpected instructions, credentials where they shouldn’t be.

Isolation — The machine is expendable. The blast radius is limited.

Let your Clawdbot set this up for you

Here’s the fun part: you can share this article with your Clawdbot and have it implement these security measures for you.

Copy and paste this prompt:

I want you to harden our security setup based on this article: [paste article URL or content]

Specifically:
1. Check if our gateway is exposed (bind setting) and fix if needed
2. Set up Bitwarden CLI for secrets management with a secure wrapper script
3. Add strict rules to SOUL.md about never displaying secrets
4. Add content quarantine / trust levels to our security rules
5. Set up git tracking for the workspace with a proper .gitignore
6. Create a weekly security audit cron job for Sunday nights that also checks https://docs.clawd.bot/gateway/security for updates
7. Add ACIP prompt injection defense rules to a SECURITY.md file
8. Set up incident logging in memory files
9. Know how to rotate sessions if credentials get exposed
10. Install LuLu (or similar) for network monitoring
11. Add soft limits / circuit breaker rules for bulk and destructive operations
12. Document everything in a Security.md file

Ask me for any permissions you need. Walk me through anything that requires my input (like unlocking Bitwarden or approving LuLu permissions).

Your Clawdbot will walk you through the setup, ask for permissions where needed, and document everything.

What would it take to reach A-tier?

I’m being honest: B+ is where we are. Here’s what’s still missing and what it would take to close the gap.

Verify disk encryption
Is FileVault enabled? If someone steals the physical machine, can they pull the drive? Run fdesetup status to check. If it’s off, turn it on.

Push git to a private remote
Right now our git history is local only. If the machine dies, we lose the rollback capability. Set up a private GitHub or GitLab repo and push regularly.

Review elevated tool permissions
We have tools.elevated: enabled in our config. What does that actually allow? I haven’t fully audited the blast radius. Someone should.

Bitwarden session file on disk
Our session token sits in ~/.secrets/bitwarden_session.sh with chmod 600. Better than plaintext config, but if the machine is compromised, that file is readable. 1Password has op run which injects secrets into a subprocess without ever writing to disk. Bitwarden doesn’t have a clean equivalent.

Consider a more secure channel
Telegram bot API traffic isn’t end-to-end encrypted. Messages go through Telegram’s servers. For truly sensitive operations, self-hosted Matrix or Signal would be better.

Proactive credential rotation
We rotate on incident, but not on a schedule. Best practice is periodic rotation regardless of whether something went wrong.

Technical enforcement, not just policy
Content quarantine and soft limits are rules I follow. They’re not technically enforced. If I hallucinate or get confused, I could still do something wrong. Real sandboxing would be better.

None of these are hard to fix. They’re just not done yet. Maybe next weekend.

Is this overkill?

I genuinely don’t know.

But Clawdbot has access to my emails, messages, and files. It can run code. It can send messages on my behalf. That’s a lot of trust to hand over.

I’d rather feel slightly paranoid than find out I wasn’t paranoid enough.

The demos are cool. The security is your responsibility.

Credits

This security setup was inspired by posts from the Clawdbot community:

• @Shaughnessy119

  — on running Clawdbot on a dedicated VPS for isolation

• @0xSammy

  — for surfacing the 923 exposed gateways and the bind: loopback fix

• A community post on secrets management with 1Password that I adapted for Bitwarden (secrets never in code, fetch on-demand, let them die with the shell)

• Another post on the content quarantine framework (trust levels, proactive disclosure, expendable infrastructure)

The Clawdbot community is figuring this out together. These are the people sharing what they learn.

If you give me 10 minutes, I’ll to teach you how to build an AI assistant that will save you a few hundred hours this year.

It’s easy to get started, and you don’t have to be nerd to figure it out. But if you are a nerd, that’s cool too.

I’ll walk you through everything step-by-step.

Want the shortcut?
I turned this entire setup into one prompt. Paste it into Claude Code and it will create the folders, files, memory, and three commands for you.
Get the Personal AI Assistant setup prompt →

This will be especially helpful if you do mouse and keyboard work, manage a to-do list, or want to be more productive.

You see, everyone is doing task management wrong.

I was too.

Tell me if you can relate to this:

At one point, I had 9 different to-do lists. Each one was overflowing with junk.

I followed all the best task management advice:

• Inbox

• This Month

• This Week

• Today

• Agenda

• Waiting for Others

• Done

• (plus 2 other personal lists)

The lists were supposed to help organize the chaos, but became extra stress. Every week I’d have pruning sessions, rewriting items, removing ones that didn’t matter anymore, updating tasks that changed. I was constantly processing, processing, processing.

As old items were purged, new ones showed up.

My task management system became background anxiety. Every time I got off a call or left a meeting, I’d know it was sitting there. Waiting for me.

Over time, I set a goal to keep my to-do list from having scroll bars.

Not to be productive.

Not to get ahead.

To not have scrollbars.

It sucked.

Then I saw a video that broke my friggin’ brain:

I was doom-scrolling X before Christmas break, and I saw a product manager who built an AI-powered task system using Claude Code. She typed a single command into the terminal on her computer, the AI thought for a moment, looked at her notes, her priorities, and spit out a structured agenda for her day.

No pruning lists, no reorganizing tasks, it just... worked.

I couldn’t believe it.

What black magic voodoo devil did she sell her soul to for this wizardry?

BECAUSE I NEED IT!

Then a second realization hit me:

I’ve been using AI through a keyhole this whole time.

Talking to Super-Intelligence Through a Tiny Window

I’ve been talking to AI through interfaces. Little text fields, chat boxes, tiny narrow openings.

Through chat, you’re the curator. You have to manually pull together information and files that you want AI to analyze. But if you can’t find the files you need, or can’t remember you have them in the first place, you’re out of luck.

The issue is that a partition exists between you and the AI. It’s as if it were in another room.

It’s like when you’re at the bank drive-through trying to pass paperwork back and forth to the teller through those pneumatic air tubes.

This is fine for a question like, “How many more days until the weekend?” but horrible for a request like, “Build me an on brand, 12 slide powerpoint deck on Project X from my meeting notes.”

Through a chat box, AI can’t scan your meeting notes from 8 months ago where you already solved this exact problem before. It can’t help you find a document you put in some random folder about your Q3 strategy. It can’t chase a thread about Project X from one file to another to give you a holistic view of progress.

This is where Claude Code comes in and changes things. When you take AI and give it direct access to your filesystem, the aperture blows wide open.

The walls come down.

Suddenly, you find yourself standing in the same room with this new alien super-intelligence.

David Allen Was Only Half Right

I remember reading David Allen’s Getting Things Done and absorbing the insight that “the mind is for having ideas, not holding them”.

The first time I built a to-do list using the GTD method was a huge relief. Suddenly, I had a place to dump every idea that came to mind. It got stuff out of my head. Closed loops. I didn’t have to burn mental energy holding ideas in place anymore because I had a system I could trust.

But it created a new kind of misery:

To-do list management.

But now, thanks to Claude Code, I don’t have to do that anymore either.

My AI Assistant frees me from the entirety of admin work. I have more mental energy to not only have ideas, but to execute on them.

Allen says the mind isn’t meant for holding ideas. He’s right.

I’d also add that the mind isn’t meant for processing, organizing, and administrating them either.

Your brain isn’t a filing cabinet.

It’s an engine.

And it wants two things:

1. To have ideas

2. To execute on them

Everything in between - the sorting, the pruning, the endless reorganizing - that’s friction.

Overhead.

The work that gets in the way of work.

And now you don’t have to do it anymore.

The LLMs that started in chat boxes can now be installed on your computer. They can manipulate files, take action, and do things in your digital life.

The super-intelligent brain is growing hands and feet.

So, what does this look like?

Let me take you on a tour.

Want the shortcut?
I turned this entire setup into one prompt. Paste it into Claude Code and it will create the folders, files, memory, and three commands for you.
Get the Personal AI Assistant setup prompt →

How My Personal AI Assistant Works: Two Notes, Three Commands

This system saves me at least 1 hour per day.

It’s simple. Two notes, two folders, one memory file, and three commands.

The notes are a scratch pad for quick capture, and a task list. The folders hold daily notes and meetings. The memory file is where Claude maintains working context over time, so it remembers what matters from one day to the next.

For the three commands, I created /start, /sync, and /wrap-up.

That’s it.

The scratch pad is where I dump everything throughout the day. Bullets, notes, decisions, follow-ups. Anything that crosses my mind. I don’t organize it. I don’t file it. I capture it.

The daily note is like a captain’s log. Decisions made, meeting links, completed tasks. A record of the day. I don’t touch this, AI takes care of it for me.

The memory file is the interesting one. Claude maintains it over time. Working context that carries forward. What I’m focused on, what’s waiting, what matters.

This is a simple starting point that anyone can use.

I’ll take you through it step-by-step.

What Happens When I Type /start, /sync, and /wrap-up

When I run /start in the morning, Claude looks at my tasks and memory. It surfaces aging items, things that have been sitting too long. It asks about completed items. It helps me set priorities and pulls up anything from memory that’s earmarked for that day.

When I run /sync during the day, Claude reads my scratchpad and updates my to-dos. It reads any meeting transcripts I’ve dropped in, summarizes them, renames the files, and updates my daily note with what happened.

When I run /wrap-up at the end of the day, Claude checks for any unprocessed meetings, reviews my to-do list and scratchpad one more time, and writes to memory for continuity. Tomorrow’s /start will pick up right where today left off.

Here’s a real example of how this helps. I had this item on my list and I couldn’t remember where it left off from a week ago. So I asked Claude. It found the meeting where it was discussed, went through the transcript and memory, and came back with the specific answer. Turns out, one of my colleagues reached out to our engineering team to get a report created. That’s it. That’s what I needed. No searching. No scrolling through notes. Just the answer.

What would have taken me ~10m of searching to find, it pulled up in seconds.

Now I work. The AI handles the rest.

How to Set Up Yours in an Afternoon

I’m going to walk you through the exact steps I used to build my own AI Assistant, so you can create your own. I’ll explain the principles and fundamentals that make this system work, so you can modify it for yourself.

By the end, you should be able to follow these steps and experience the same benefits I have.

Let’s get into it.

What you need:

• Claude Code installed on your machine

• A notes app (I use Obsidian, but any app where notes are stored as files works)

• A basic folder structure

Step 1: Install Obsidian

Obsidian is a free notes app that stores everything as plain text files on your computer. This is important, so that Claude can read and edit those files directly.

1. Go to obsidian.md

2. Download the app for your operating system (Mac, Windows, or Linux)

3. Run the installer

When you first open Obsidian, it’ll ask you to create or open a “vault.” A vault is a folder on your computer where your notes live.

1. Click “Create new vault”

2. Give it a name (I called mine “Work”)

3. Choose where to save it — somewhere easy to find, like your Documents folder

That’s it. You now have a notes app where everything is stored as files Claude can access.

Step 2: Install Claude Code

Claude Code is a command-line tool that lets Claude work directly with files on your computer. This is the key that unlocks everything.

On Mac:

1. Open Terminal (press Cmd + Space, type “Terminal”, hit Enter)

2. Copy and paste this command, then hit Enter:

npm install -g @anthropic-ai/claude-code

On Windows:

1. Open Command Prompt or PowerShell

2. Run the same command:

npm install -g @anthropic-ai/claude-code

Don’t have npm?

If you get an error about npm not being found, you need to install Node.js first:

1. Go to nodejs.org

2. Download the LTS version

3. Run the installer

4. Close and reopen your terminal

5. Try the npm install command again

Verify it worked:

Type this in your terminal:

claude --version

If you see a version number, you’re good.

Step 3: Create your folder structure

Now let’s set up the folders inside your Obsidian vault. Here’s what the structure looks like:

In Obsidian:

1. Right-click in the left sidebar → “New folder” → name it “Daily Notes”

2. Right-click again → “New folder” → name it “Meetings”

3. Right-click → “New note” → name it “Scratch Pad”

4. Right-click → “New note” → name it “Task Board”

The .claude folder and memory.md file will be created automatically when you set up your commands.

Step 4: Connect Claude Code to your vault

Now we connect the two pieces. This is where we’ll spend a little time in the terminal.

What is terminal?

Terminal is a way to talk to your computer by typing commands instead of clicking around. It looks like something from an 80s hacker movie, but it’s not as scary as it seems. You type a command, hit Enter, and your computer does something. That’s it.

You only need two commands to make this work:

That’s the whole vocabulary. Two words.

If you get stuck: Your favorite LLM (ChatGPT, Claude, whatever) or Google are your friends here. “How do I navigate to a folder in terminal” will get you unstuck fast. Don’t be afraid to ask.

Navigating to your vault

Now you need to tell terminal where your Obsidian vault lives. There are a few ways to do this:

Option A: The shortcut (easiest)

Skip the cd command entirely. Open terminal directly inside your vault folder:

On Mac:

1. Open Finder and navigate to your vault folder

2. Right-click on the folder

3. Choose “New Terminal at Folder”

On Windows:

1. Open File Explorer and navigate to your vault folder

2. Right-click in the empty space inside the folder

3. Choose “Open in Terminal”

Done. Terminal opens and you’re already in the right place.

Option B: The cd command

If the shortcut isn’t available, you can navigate manually using the cd (change directory) command.

Type cd followed by the path to your vault folder:

cd /path/to/your/vault

For example, if your vault is in Documents and called “Work”:

cd ~/Documents/Work

Pro tip: You can also type cd (with a space after it) and then drag your vault folder from Finder/Explorer directly into the terminal window. It’ll paste the path for you.

Hit Enter. You won’t see much happen, and that’s normal. No news is good news in terminal.

Starting Claude Code

Now type:

claude

And hit Enter.

Claude Code is now running and can see all your notes.

A note on what Claude can see: Whatever folder you launch Claude in, it can see everything in that folder and all subfolders. That’s what makes this powerful, but it also means you should be thoughtful about where you run it. Don’t launch Claude from a folder containing passwords, API keys, or sensitive documents you don’t want processed.

Take it for a spin

Before we set up commands, talk to it. This is the same Claude you’d use on claude.ai, you’re just talking to it in terminal instead of a chat window.

Try a few things:

• “What files and folders do you see in this directory?”

• “Summarize what’s in my Task Board note”

• “Search the web for today’s top news headlines”

• “What’s the weather in [your city]?”

Get a feel for it. Ask it questions. Give it a task. The more you play around, the less intimidating it feels. Once you’re comfortable, we’ll set up the commands that turn this into a real system.

Feeling ambitious? You can also use this prompt to have Claude build all three commands at once.

Step 5: Set up your commands

This is where the magic happens. You’ll create three commands that tell Claude what to do.

Here’s the good news: you don’t need to figure this out yourself. You’re going to ask Claude to build these for you.

I told you no tech skills were required.

How commands work

Commands are text files that live in a special folder. When you type /start in Claude Code, it looks for a file called start.md in the .claude/commands/ folder and follows the instructions inside.

I like to think of these like prompt shortcuts. I have no idea if that’s what they actually are, but that’s what it seems like.

Creating your commands

In your Claude Code terminal, copy and paste these prompts one at a time. Claude will create each command file for you.

For /start (beginning of day):

I want to create a /start command that I run at the beginning of each workday.

When I run /start, you should:
1. Read my Task Board.md file and summarize what's on my plate
2. Read the .claude/memory.md file to check for any context or follow-ups
3. Look at my calendar or any time-sensitive items
4. Help me prioritize and plan my day
5. Ask if there's anything from yesterday I completed that should be removed

Create the command file at .claude/commands/start.md

Keep the instructions clear and concise. I want this to feel like a quick morning standup with an assistant.

Claude will create the folder structure and the command file for you. It might ask a few clarifying questions — answer them however feels right for your workflow.

For /sync (throughout the day):

I want to create a /sync command that I run throughout the day to process my notes.

When I run /sync, you should:
1. Read my Scratch Pad.md file and process any notes I've captured
2. Check the Meetings/ folder for any new transcripts and summarize them
3. Update my Task Board.md with any new to-dos that came up
4. Update or create today's daily note in Daily Notes/ with a summary of what happened
5. Clear out the Scratch Pad once everything is processed

Create the command file at .claude/commands/sync.md

I want this to feel like handing a pile of loose notes to an assistant and having them file everything properly.

For /wrap-up (end of day):

I want to create a /wrap-up command that I run at the end of each workday.

When I run /wrap-up, you should:
1. Check if there are any unprocessed meetings or notes in Scratch Pad
2. Do a final update to my Task Board.md
3. Update today's daily note with a summary of the day
4. Write anything important to .claude/memory.md so you remember it tomorrow
5. Give me a brief "end of day" summary

Create the command file at .claude/commands/wrap-up.md

I want this to feel like closing out the day cleanly — nothing falling through the cracks.

When you’re finished, test each command in sequence to make sure they work. If you run into any issues, just ask Claude to help you fix it.

What you now have

These are starting points. As you use them, you’ll want to tweak things. Maybe you want /start to pull in your calendar. Maybe /sync should also check your email folder. Ask Claude to update the command, it’s all text files.

The to learn is through play. Follow your curiosity. Try stuff. As long as you monitor what Claude’s doing, it’s low risk.

Step 6: Set up memory

Memory is what gives Claude continuity between sessions. Without it, every time you start Claude Code, it’s a blank slate. It’ll have no idea what you talked about yesterday, what’s on your plate, or what you’re waiting on.

The memory file fixes that. It’s a simple text file where Claude writes down important context: ongoing projects, people you’re working with, things to follow up on, decisions that were made. Claude reads it at the start of each session and updates it at the end.

Let’s have Claude create it. Paste this:

I want to set up a memory file that you'll use to maintain context between our sessions.

Create a file at .claude/memory.md with a simple structure for:
- Current projects or priorities I'm focused on
- People I'm working with and any relevant context
- Things to follow up on
- Recent decisions or important context

Keep it concise — this should be a quick-reference file, not a journal. You'll read this at the start of each session and update it during /wrap-up.

For now, create the structure. We'll fill it in as we work together.

Claude will create the file with a clean structure. Over time, as you use /start and /wrap-up, this file will become a living document that keeps Claude up to speed on your world.

Step 7: Test it out

Let’s make sure everything works.

1. Open Terminal and navigate to your vault (or wherever your vault lives):

cd ~/Documents/Work

2. Start Claude Code:

claude

3. Run your first command:

/start

See what it surfaces. Add some notes to your scratch pad, drop in a meeting transcript, run /sync. Watch it work.

The complete picture

Here’s everything you built:

Want the shortcut?
I turned this entire setup into one prompt. Paste it into Claude Code and it will create the folders, files, memory, and three commands for you.
Get the Personal AI Assistant setup prompt →

But What About...

If you made it this far and you’re new to this, I’m sure you might have a few concerns.

Let’s see if I can help.

“I’m not technical enough for this.”

Look, I get it. The terminal looks intimidating. There’s a learning curve. But it’s 3 commands, a scratch pad, some folders. No coding required. Besides that, you’re talking to a super-nerd that is endlessly patient. If you can order a pizza over the phone, you can figure this out.

If you want a more visual starting point, Claude Cowork is the easier on-ramp.

“AI will mess up my files.”

Claude Code asks permissions before making changes and you’ll get a chance to review and approve everything. It’s more like a fast assistant handing you drafts than a robot with unsupervised access. Start with non-critical files first if you’re nervous. And never let it do something if you’re unsure what it’s asking.

“What about privacy?”

Honest answer: data does go to the cloud. You can set it not to train on your data, and it’s removed within 30 days. I’m careful what I share. Concern level for me is about a 2 out of 10. My eyes are open, but I’m not worried.

Six Months From Now

The future of knowledge work is already here, but most people are six months behind.

I don’t say that to be dramatic, it’s just true.

I don’t have nine Trello lists anymore. I’ve said goodbye to endless pruning, processing, and manual admin. There is no more background anxiety.

When I finish a meeting, I paste the transcript and forget about it. When something pops into my head, it goes on the scratch pad and I move on. All I have to do is type /start in the morning and Claude tells me what matters.

Now, I just work.

People who figure out who to work alongside AI for brainstorming, processing, organization, prioritization, note taking, and more - not through chat boxes but directly in their filesystem - are going to operate on a different level.

The capability already exists.

Start using it. Break things. Follow your curiosity. The system I showed you is the starting point, not the final destination.

We’re in pioneering territory. There are untold use cases we haven’t even discovered yet. Knowledge work is changing in front of our eyes and soon everyone will feel it.

The door is open. Go see what’s possible.