A month ago that same deck would have taken me two days of arguing with Claude about fonts and spacing 😭 I have a design background (which made it worse) and I've tried everything on this.

Nothing worked until I stopped describing the deck I wanted and pointed at one I already had.

In this issue:

• What the AI gives you when it has nothing of yours to read

• How to connect Figma to Claude, step by step

• The four files that make everything come out on brand

If you are totally new to Claude Code, I cover 80% of it here in 20 minutes.

First, what I (actually) tried

It was Thursday and I needed a deck in under 24 hours. I already had one created by Claude Code. It was pretty good, but I needed something bespoke for tomorrow.

I was tired of fighting Claude. So I gave the same brief to the tools everyone recommends.

Here is Gamma.

Here is Claude Design.

Here is Canva, through its own MCP connector (which I'd never properly used before).

All of them are competent. None of them were perfect.

Here is the same slide out of my own system.

It would have taken me hours of voice-to-text to perfect the screenshots, the layout, and everything about these presentations.

You cannot fix that by prompting harder. They had nothing of yours to read, so they had nothing to point at.

I’m not bashing Gamma, Claude Design or Canva here.

I’m just simply saying I found a better solution.

What (actually) worked

Last year I worked with Shahmeer on a presentation. He built the first draft properly, the way a designer does, and it sat in my Figma for months.

I had never once thought of it as a system. I just liked the deck.

You almost certainly have a file like that (everyone does). You've just never called it the reference.

The Figma MCP is what turned it into one.

Claude reads a real Figma file directly now, so instead of describing my brand I pointed at that presentation and dropped in the ideas I wanted to make.

Setting up Figma (properly)

1. Make a Figma account. Go to figma.com, sign up, and take the free plan (there is a paid one, ignore it).

If you already have an account then skip this. Plenty of people do and have just never opened it.

2. Get the file you like into it. This is your reference.

It is the deck, or the brand board, or the one thing somebody made you that actually looked right (mine was a deck).

If a designer built it, message them and ask for the Figma link. They will still have it.

Mine sat in my account for months and I had never once opened it.

3. Connect Figma to Claude. In the Claude desktop app, click the + button in the bottom left of the chat box.

Hover over Connectors, then click Manage connectors.

Hit the + next to Connectors, find Figma in the list, and click it.

A browser window opens, you log into Figma, and you approve the access.

4. Copy the link to the exact thing you want. This is the step everyone misses, and I missed it too.

Do not grab the web address out of the top of the browser.

Open the file in Figma, right click the frame or the slide it needs to read, and choose Copy link.

That link points at that one thing instead of the entire file.

And that is the difference between Claude reading your design and Claude reading a filing cabinet.

5. Send it, with the thing you are stuck on. Paste the link first, then attach whatever you are currently fighting with, then say what is wrong with it and what you want instead.

What I typed was closer to a confession than a prompt.

That was it. Claude read the real file, took the styling out of it, and rebuilt the deck in it.

I couldn’t believe it, now I could update to the presentation whilst keeping it perfectly on brand just by describing what I wanted to Claude.

It matched my brand style exactly, because it was built on top of something that already did.

Why my infographics run somewhere else

I run my LinkedIn graphics on a similar system I’ve spent months building. It goes above and beyond Figma MCP, Claude Design and every tool I have tested, and it’s not even close.

I tested the Figma MCP for infographics (it’s pretty good), but it struggled with the complexity that my infographics require.

My Claude Code system is far superior.

It is that good for one reason. I worked with Nick Broekema, and Nick wrote me a proper brand guideline. He gave me the hex codes, the type scale, the spacing, everything.

I have built everything since on that one document. You can get to the same place with yours. Start with the reference. It is the thing everything else stands on, and it will change your output this week.

I've spent months on mine, and that whole system (the references, the templates, and the checks that stop a bad graphic ever reaching me) is what I'm opening up as the LinkedIn AI OS. Waitlist here: growth.charliehills.io/are-you-a-fit

Step 1. Pull your rules out of the work you already have

If you can pay someone to design, pay them once, not per graphic.

And if you can't, pull it out yourself.

Go and collect five pieces of design you already like. Your best graphics, your banner, your website, a deck you reuse, anything of yours that felt right. Screenshots are fine.

Put them in a folder (that's what the prompt below reads).

Paste this:

I am giving you five pieces of design I already like.

Write me one file called REFERENCE.md that any AI can read before it makes anything for me. It must cover:

1. Every colour as a hex code, and what each one is for
2. The fonts, the sizes, and the spacing between things
3. Where my logo goes and how much clear space it needs
4. Five things my brand must never do
5. One example, described in full, of it done right

Ask me about anything you cannot work out from the files.

Show me the file before you save it.

Run this once and you keep the file forever.

So the next test was a brand that is not mine.

Tarana Kasana is a really good LinkedIn creator (go and follow her). She did not ask me to do this, I just wanted to see whether the system held up on a brand that was not mine, so I built hers off her public LinkedIn banner and nothing else.

Perfectly on brand, and not one value was a guess. Your banner is probably already your reference (you have just never read it as one).

Same day, same system, completely different brand. Jim Cranston did not ask me either, and his came out in his own sage green with nothing of mine anywhere in it.

That is how both came out so fast. Not a clever prompt (it never is). A library of work to point at, there’s sixty-six builds in my system now, so it clones the closest one instead of starting again.

You can build that library yourself, one graphic at a time. It just takes months, and mine took most of a year. If you want my full system and templates rather than months of building your own: growth.charliehills.io/are-you-a-fit

Step 2. Make Claude read it before it draws

A reference nobody opens is a mood board with extra steps.

Paste this:

Add this to CLAUDE.md, and create the file if it does not exist:

Before designing, generating or laying out anything, read REFERENCE.md in full.

Every colour, size and spacing value comes from that file.

If something I ask for is not covered there, ask me rather than choosing for yourself.

When you have finished, check your own output against REFERENCE.md, fix what fails, and only then show me.

That last line does more than the other three combined (by a distance). Claude checks its own work against the file before anything reaches you.

Step 3. Turn the file into a design system

The reference tells Claude what's allowed. A design system tells it how to build (very different jobs).

Paste this:

Read REFERENCE.md and every file in /examples.

Write me DESIGN.md in the project root. It is the build rulebook, not a mood board. Cover:

1. The colour tokens, named by JOB (background, ink, accent, muted), each with its hex
2. The type scale, in real sizes, and which face goes where
3. The spacing scale, as a small set of values I reuse everywhere
4. The components I actually use, with what each one is for
5. A decision log at the bottom, where every choice we make from now on gets recorded

Work from what is IN the reference and the examples. Where they disagree, ask me.

Show me the file before you save it.

Then give it somewhere to live, because a rulebook nobody can find is a rulebook nobody reads:

your-project/
  CLAUDE.md        the four lines that make it read everything below
  REFERENCE.md     the five rules
  DESIGN.md        the build system
  examples/        five things you were happy with

That folder is the whole thing. Four files, one of them four lines long.

Screenshots in a folder are inspiration. A file it reads first is a system.

And never accept the first version. Ask for three or four variants and put them side by side, or you'll end up polishing whatever came out first (I did this for months).

The 29 tells of AI design

I can only spot these because I have been at this a while. You probably can’t yet, and every one of them came out of a graphic I rejected.

I've built a free skill off the back of six months designing in Claude Code. It checks your graphic against all 29 before you post it.

I've written before that graphic design was dead. That’s half right.

AI has replaced the designer who makes the individual post. It has not replaced the person who decides what good looks like. That judgement was never in the file, and it never gets automated.

It came from two people. Shahmeer made the deck. Nick made the guideline. I have not paid either of them a second time, and both of them are still in everything I ship.

A machine can hold your rules perfectly. It cannot tell you they were the right ones.

This Thursday from 4pm - 5pm (UK). I’m hosting LinkedIn Summer School #3 (my final free workshop of the summer) and I’m running the whole thing live.

I’ll have multiple Claude Code terminals open, so you can watch the reference file do the work instead of taking my word for it. luma.com/linkedin-olrf

Then AI Summit Barcelona, 22 and 23 September, where I am running a workshop on how I (actually) grew to ~250k on LinkedIn. If you cannot make it to Barcelona, the Digital Pass gets you my workshop plus the full programme, live and then on demand for a year.

You also get 10+ hours of hands-on training, credits for Anthropic, OpenAI, ElevenLabs and Zapier, an official certificate and a LinkedIn skills badge. Passes and details: aisummitbarcelona.com

That's it from me this week. See you again next week.

Stay curious, stay human and keep creating

— Charlie

P.S. What did you think of this edition? Reply and tell me, I read every one.

I'm not a developer. I'm a marketer who discovered the power of Claude Code for content. I might sound strange to say that, because coding is for developers, right? Wrong.

Coding is for everyone.

You've heard the term "vibe coding" before, I'm sure. The practice of describing what you want to AI and it'll build whatever you want. The same applies to marketing.

Every day I'm "vibe designing" or "vibe marketing" with Claude.

I've spent the last 173 days mastering the practice, so in this article I hope to offer you a glimpse of what it has done for me and what it could do for you, too.

This covers everything I have learned in 173 days.

Anthropic just dropped three new features worth knowing about if you are running Claude Code properly.

If you’re already comfortable in Claude Code, start here.

If you’re still learning, read the article, then watch :)

Before you scroll: if any of this is useful, someone on your team has the same problem and doesn’t know it.

First, what this thing actually is

Claude Code used to terrify me.

1. I'm a marketer by trade.

2. I'm not (actually) technical.

3. I'm definitely not a developer.

So I thought, why should I be using Claude Code?

Well, Claude Code has changed the way that I run my content business. And it's not just for building applications. You can use Claude Code to create content, too.

It reads, writes and edits your files. It calls APIs. It browses the web.

It spawns workers to run jobs in parallel. It runs scripts you never wrote.

You do not need to be a developer. The barrier around tools like this is largely gone, and a few hours of learning-by-doing gets you real results.

#1 Get it running (about two minutes)

Day 1. I opened the terminal and typed “claude” with no idea what I was doing. Nothing I built in the first month compounded. Every session started from nothing and I re-explained myself.

Get a paid Claude plan first. Pro at $20/month is enough to start. Max gives you more usage if you end up running heavy workflows.

Install it with one command in your terminal, from code.claude.com/docs.

Takes about two minutes.

If you've never opened it, you access Claude Code through the terminal or through the Claude desktop app. I run mine through the terminal.

The desktop app is pretty good. Start there if you’re new.

Type claude. It prompts you to log in.

Now navigate to a folder you actually work in. You're running.

What you need to build out here is a CLAUDE.md file. This is basically your background on who you are, what you like, and how you exactly want Claude Code to work.

One of the best practices here is to simply type this:

/init

This will initialise a new CLAUDE.md file for you.

But you don't have to follow this step. To be honest, you can just talk naturally to Claude Code, and then ask it to update your MD file in retrospect, based on the learnings that you make along the way.

That second way is the one I'd actually recommend, because you don't yet know what's worth writing down.

#1.5 Then open /config

Don’t judge me, it’s day 173, and I only discovered this panel today.

/config is where Claude Code keeps its settings. About forty of them.

I'm not going to pretend I've tuned them. Most are still on whatever they came with. The point is that on day 173, writing a guide about this thing, there was a whole panel I'd never opened.

That's the thing about Claude Code. You find something new and work out how to use it every single day. With Claude Chat and Cowork you're fairly limited in what you can actually do. The ceiling is (almost) uncapped.

#2 Plan before it builds

You can slash command /plan, and Claude Code will build out a plan before it builds anything. I'd highly recommend using that command if you're totally new to Claude Code, like I was in March.

There are two ways in. Type /plan, or hit shift + tab. Same mode, and shift + tab cycles straight back out again.

Plan mode is read-only, and it is far cheaper than rebuilding. A failed build costs more than ten minutes of planning ever will.

This is the alignment step that runs before Claude builds anything. Claude proposes a layout, asks me what I want changed, and then it codes.

Pro tip: if you're mid-task and the output starts feeling off, drop back into plan mode to realign on where you actually want to take it. This is way better than letting Claude run down the wrong path and burning tokens.

#2.5 Auto mode

When I first started using Claude Code you had to bypass permissions, because otherwise it asks you a tonne of questions and it is genuinely annoying.

That is probably the first thing that blocks you when you get started. Not the terminal or the setup. Just being interrupted every thirty seconds.

That is what auto mode is for, and it is now the default. It is safe to use, and honestly, as long as you are not doing anything too sensitive, you are good.

The reason I am comfortable saying that is what sits underneath it. In auto mode every thing gets checked before it runs, and it deliberately ignores your own allow-list rather than trusting it. Shift + tab cycles you between manual, auto mode, and plan mode.

#3 Install things

This is the section I’d move to day one if I could start again.

I spent my first month building my own versions of things that already existed, were better, and were free to install.

Type this:

/plugin

That opens the marketplace.

1. Superpowers

To be honest, I only use this for brainstorming, and I do not even use that much.

/plugin install superpowers@claude-plugins-official

2. gstack

I use this for one thing. The agent browser.

It opens a real browser and reads the live web, so you get what is on the page today.

You find it on GitHub, click code on the top right, and copy the installation command.

github.com/garrytan/gstack

My favourite thing about this workflow is that I built a system that writes my newsletter with me. Claude Code can now do a version of this itself (without gstack).

claude --chrome

It opens tabs in your own Chrome, the one already on your desktop, and it carries your logins with it. So it can work inside Gmail, Google Docs, Notion, anything you are signed into, with no connector set up at all.

3. HyperFrames

Writes HTML and renders it as video. I use it for the HTML video editing.

Almost every YouTube video I ship comes out of this. I describe the overlay, Claude writes the HTML, it comes back as an mp4.

/plugin install hyperframes@claude-plugins-official

4. Claude HUD

This one sits at the bottom of my screen every time, so I can see my context and my usage.

What it’s doing, and what it’s spent.

/plugin marketplace add jarrodwatts/claude-hud
/plugin install claude-hud@claude-hud

5. Caveman

I installed this on day 72 to get short replies.

It gives no preamble or summary, just the answer.

/plugin marketplace add JuliusBrussee/caveman
/plugin install caveman@caveman

Pick the one for your job

Those five are mine because of the job I do. Yours will be different.

If you want somewhere to start that is not my list, Matt Pocock has a skills repo that is going up fast. 224k stars: github.com/mattpocock/skills

Which is the exact problem I got tired of. Every week someone asks me which repo to install, and whether you install it depends on what they do all day.

So I built the Claude Code GitHub Repo Bible.

There are 268 repos worth installing, in seven chapters, every link checked live and ranked by stars.

If you’re already a subscriber, it’s sitting in your resource vault right now.

If you’re new here, subscribe and it opens up below.

#4 Connect tools

All your connectors carry over from claude.ai. You don’t set them up again.

And I only keep the tools I genuinely use every day.

Connectors are one of the most useful features in Claude.
Connect Gmail and your calendar, and Claude can read both.

Ask it for a summary of your week and it actually knows what is in there. Then it can act inside those apps on your behalf.

To add one, go to customise, then connectors. There is a list of pre-built ones to browse.

Configure them in the web interface, and your Claude Code sessions inherit them.

Or you can add custom connectors, like I’ve done.

Google the tool name plus "MCP". Most things have a connector by now.

If yours doesn’t, use Zapier. It carries 9,000+ integrations, so whatever you run is in there somewhere.

There’s really no excuse anymore.

Worth two minutes on day one, because you are probably already holding more than you think:

List every tool you can reach right now through my connectors. For each one, tell me what you could do with it today without me setting anything else up.

I don’t leave the terminal, every tool gets called from there.

#5 Brief it

Honestly, I didn’t really start with a CLAUDE.md.

I started with Boris Cherny’s, because it is the best starting point there is. He built Claude Code. I covered that here: Claude Code (for normal people).

CLAUDE.md is the single most important file in your entire setup.

Every project folder, skill and workflow you build inherits one of these.

It’s markdown, which is just text. It’s basically a system prompt. That’s all it is. It’s nothing complex. Mine holds my brand rules, my voice and my output format in one place.

Write your CLAUDE.md before you write a single prompt, and everything else adds on top of it. Where I went wrong was what I thought it was for. I treated it as the place everything lives. All the rules, all the context, every correction.

It’s your front desk, not your filing cabinet.

Then you point Claude at a specific folder. This is separate from your global one.

Global CLAUDE.md is the instructions that apply everywhere.
Local CLAUDE.md is the instructions for one project.

It’s basically the same as your custom instructions in ChatGPT or Claude, and your project instructions in the chat interface. Same idea, just sitting in a file instead of a settings panel.

Context is what keeps the two from bleeding into one another.

Keep them apart and neither one bloats.

Mix them and your global file becomes a landfill. That is exactly what mine became, because every correction feels global at the moment you make it.

Almost none of them are.

#6 Folders

I have one folder per social platform. That is my system.

LinkedIn got its folder on day 24. YouTube on day 65. Instagram and Substack on day 93. It took me three months to cover four platforms, one at a time, as each one got busy enough to need its own rules.

The restraint is the useful half. Do not open the second folder until the first one is genuinely busy enough to need its own rules.

I have been slowly rolling out folders as I build out my content ecosystem.

You don’t need to set up folders on day one. Perfect the most important thing first, which in my case was LinkedIn. Then gradually roll out folders, skills and MD files to the rest of the system.

LinkedIn holds 1,586 files and 1,037 lines of rules. X is three files.

A file that only loads for one job can be specific in a way a global file never can.

Never have one giant file.

Paste this, inside a project folder:

Read this project and write a CLAUDE.md for it. Include the commands I actually run here, my style, and the three mistakes you are most likely to make in this folder.

I (finally) wrote my Instagram project’s CLAUDE.md writing this newsletter :)

#7 Memory

If you don’t have memory enabled, this is the simplest fix you can make to your Claude experience.

Memory files I wrote per month: March 11, April 48, May 82, June 129. Then July, 590. You see the sharp rise in my memory. It begins the month after I started writing sessions down properly.

Just tell Claude to update its memory and it will.

It also does a chunk of this on its own now. Auto memory means Claude writes down what it learns as it works, across sessions, without you asking. Just type /memory to enable it.

Here’s example of it in action:

This makes the folder job smaller than it was when I started.

You are steering what it keeps, not typing it all out yourself.

You don’t need a complex memory system. You don’t need to learn graph engineering. You just need folders, made the way you want them to work.

Say you are doing design work. You need your references and your brand guidelines. You just need to subcategorise them.

Paste this:

Set up a memory folder for me. Ask me what work I do most, then create the subfolders that work needs, and tell me what goes in each one.

Then add this one, so it keeps itself up to date:

Every time I correct you or tell you something, save it into a separate MD file. That is your memory.

The more Claude knows about you, the less you have to re-explain yourself.

Claude will make mistakes. It takes time to ramp up, but you’re building a system.

#8 Skills

Anything you've retyped twice is already a skill.

Not the cleverest thing you do. The most repeated one.

Skills are how I turn the repetitive work in my business into one-line commands I fire from any chat. They're how I take a 200-word prompt and turn it into an eight-character command.

The only way I make new ones is this. Chat with Claude until the workflow is dialled in, then ask it to save the thing as a slash command.

So paste this, right after you've had a Claude that worked:

Turn the chat above into a reusable skill. Name it, and tell me the one word that fires it.

Choose the trigger word deliberately. Too common and it fires in the middle of unrelated work, which is genuinely worse than it never firing at all.

I can’t remember the first skill I made. It is just something you intuitively do as you are doing a task, and then you ask Claude to make it into a skill. If you spend enough time and perfect a process with AI, make it a skill. If you do it twice, make it a skill.

There are three ways to get one, and none of them involve writing a file yourself.

1. Get to the end of something that worked and say, save this as a skill.

2. Or give Claude a lot of context and say, I need a skill that does this.

3. Or record one. I walked through that here: Record a skill in Claude.

#8.5 The ones I (actually) type

A command ships with Claude Code. A skill is a file you made. Both start with a slash, which is the entire reason people mix them up.

I went back through every prompt I have typed since January. Twenty-eight thousand of them, across two hundred and seventeen days. It is not the list I would have given you from memory.

Three built-in ones I genuinely reach for:

/plan
/goal
/clear

/plan is the only thing that survived from day one. I typed it at 16:57 on my first afternoon, and again this morning.

/goal I use most widely, a hundred and eight times across thirty-eight days. It goes at the end of a prompt, not the start:

You should keep going until this is fixed. I don’t want you stopping. /goal

One more, for anything I need properly researched:

/deep-research

It sends five searches out at once, then argues with itself, and anything it can refute twice gets cut before you see it. What comes back is a short report where everything left standing survived a fight, with the sources next to it.

Then the three I built myself, which I use every working day:

/save-progress
/close
/show-me

Save progress writes what the session learned into my files.

Close finds every loose end and either finishes it or hands it back as one decision.

Show me puts whatever I have just made in front of me in the browser, instead of describing it.

I’ve used /save-progress 103 times. /close, 80. /show-me 131 in the 11 days since I built it. More than any command that ships out the box.

That is the argument for building your own. The ones you make for your own friction get used harder than anything that arrives in the product.

All three are free and public: save-progress, close and show-me.

And the one I cannot give you:

/infographic-generator

It runs my LinkedIn graphics from idea to finished render. The only thing I have built and will not publish, because it is becoming a product.

If you want it when it goes live, sign-up here: https://growth.charliehills.io/are-you-a-fit

And keep them small

Don’t have a monster skill. If a skill is really comprehensive, Claude is not going to read all of it. What you want is a CLAUDE.md that directs to a skill, and that skill directs to the next one, and it cascades. So every skill is a specialist, not an encyclopedia.

I have 214 on this machine. The average one is 539 lines, and the biggest is 3,153. That one loads its whole body every time it fires, whether I needed the last 2,000 lines or not.

Paste this:

Two things, and do not change anything yet.

1. List any skill of mine doing more than one job, and tell me exactly where the split is.

2. Look back over my history and tell me which of my skills I have never once invoked. Rank that list by how long ago I made them.

The other half of that is the ones you never fire.

Of the 214 sitting on this machine, only 72 have ever actually fired. The other 142 have never run once.

I pruned mine properly a fortnight ago and wrote up exactly how in Delete your CLAUDE.md. Same job, same method, whether it is instructions or skills.

#9 Make it fire without you

Rule of thumb: CLAUDE.md guides. A hook forces.

I wrote six hooks in my first 150 days and nine in the last 23. I spent five months writing requests and wondering why they only worked most of the time.

Everything in your CLAUDE.md is a request. Claude reads it, weighs it against everything else you wrote, and usually complies.

Usually.

A hook is a script that runs automatically at a fixed moment, and Claude doesn't get a vote. Straight from the docs: to block an action regardless of what Claude decides, use a PreToolUse hook.

The test: would you be annoyed if this got ignored once?

If yes, it shouldn't be a sentence in a file.

The word is the hook. What it goes and gets is the skill.

The simplest one you could build:

When I say the word “LinkedIn”, write me a LinkedIn post.

That’s a hook.

I never tell it to. I just say the word/s.

This is where hooks and skills meet. The word is the hook. What it goes and gets is the skill. You cannot see that pairing anywhere in Claude, which is why it took me so long to work out I already had loads of them.

Start by finding out what you already repeat. Paste this:

Help me work out what I should turn into a hook.

Look back over our recent conversations and find:

1. Things I ask for over and over, in roughly the same words.
2. Times I had to tell you something you should already have known.
3. Times you did something I had to correct more than once.

For each one, tell me:
- the word or phrase I actually say when I want it
- what you should go and read, or do, the moment I say it
- whether that word is unique enough that it won't fire by accident

Give me your top three, ranked, with the best one first. Do not build anything yet.

It read 15,109 prompts I had typed over 45 days. My most repeated phrase was “show me”, 424 times, and nothing was firing on it.

Then just tell Claude to turn the top one into a hook (no special command or config needed). You describe it, it builds it, and it shows you before it saves anything.

Worth adding this to your CLAUDE.md while you’re there:

If I ask you for the same thing three times in a session, tell me, and suggest it as a hook.

#10 Routines

Up until now, if you wanted to do anything in Claude Code or Cowork, your computer had to be on. Routines fix that. They run on Anthropic’s cloud, so you no longer need to leave your machine on overnight.

You set one up from the web, the desktop app, or straight from the terminal:

/schedule

I have seventeen things running on a schedule right now. Every one of them sits on this laptop, so if the lid is shut, nothing happens. That is the gap routines close.

Me personally, I don’t like to run routines that often, because I do a lot of content work. That needs a lot of human oversight, judgment and discernment.

So anything like reporting, anything across your business which is mundane or manual, easy. Content, less so.

#11 Handover

Claude Code counts tokens, not messages. A single file costs about 800 tokens. Point it at a folder instead and you have just spent 12,000.

And a long context does not get read evenly. Models read the beginning and the end carefully, and they skim the middle. So I write a handover at 50% context, because that is the point where it starts getting stupid.

There are four moments I do it.

One. At 50% context, every time, before the session degrades.

Two. When I hit full usage on my 20x Max plan. All the terminals are open, so I write the handover before anything closes.

Three. When I log out of one Claude account and into another. You lose your connector access when you do that, so the handover carries the context across.

Four. When a session is just going wrong. The new one picks up from where the old one got to, instead of from nothing.

The prompt is this:

Write HANDOVER.md. What we did, what is left, where the files are, and the three things you would tell the next session.

Then I clear, open a fresh session, and point it at that file.

What I'd do differently

Three things, honestly.

I’d borrow first. My first plugins went in on days 36 and 46. They should have gone in during week one, before I wrote a single thing of my own.

I’d have deleted sooner. Months of instructions written one bad session at a time, and I assumed all of it was still helping.

I’d have built less. I’ve built a lot of skills and I use a fraction of them. Yours will go the same way unless you check first.

The pattern in all three is the same.

I added faster than I checked, and you have the chance not to.

This is only 80% of Claude Code. Dynamic workflows, GitHub, PRs, subagents and agent teams are all things I could have included and didn’t. You don’t need the full suite. You need to install it, use it, and work it out.

Everyone is working it out. Nobody truly knows what’s going on, honestly. The ones who put the hours in are just further along.

I’ve certainly benefited from it.

One more thing, and I only checked it tonight while writing this.

My machine says I first installed Claude Code in August 2025. A full year ago. I opened it, did not get it, and barely touched it until the spring.

So when I say 173 days, that's 173 days of using it every single day. The year before that was me having it installed and not really getting it.

Which is the whole point. It was sat there the entire time.

Personal branding, even when the subject is AI, is just documenting yourself working something out. This year that thing was Claude Code, and it has been the most fun I’ve had with AI in two and a bit years of publishing.

If you made it this far

Thank you for reading.

I have turned this whole setup into one command you can run yourself.

git clone https://github.com/charlie947/start.git
cp -r start/skills/start ~/.claude/skills/

Then type /start in Claude Code.

It asks you six questions, writes your CLAUDE.md in your own words, makes your first project folder and turns memory on. Then it reads your own prompt history, finds the phrases you retype and the corrections you keep giving, and builds those into your first skill and your first hook.

If you are brand new and have no history yet, it builds them from your answers instead.

After that it installs the three plugins worth having on day one, shows you auto mode, gives you a handover prompt, sets up one routine and connects one tool.

One of each. Not six folders, not twelve plugins, not a skill library you will never open. That is the whole lesson from 173 days, packaged so you do not have to learn it the slow way.

Everything else I have made on this

The sections above are the setup. These are what you do once it is running.

The videos:

1. How to Make AI Writing Sound 100% Human, 9:06

2. Google Gemini 3 Pro Replaces Design Skills, 9:23

3. Did Claude Code Just Replace Video Editors?, 13:56

4. How I Got 100M Views with Claude Cowork + Code, 16:21

5. Why I Replaced Canva and Figma with Claude Code, 12:40

6. Anthropic Just Dropped Claude Design (Is It a Lovable Killer?), 9:03

7. 9 Steps to Build Your First Premium Website From Scratch with Claude Code, 6:12

And every guide I have written:

1. I broke up with ChatGPT (it's over)

2. I fired my team and hired Claude Opus 4.6

3. I rated every AI release of 2026

4. Claude Code (Beginner → Advanced)

5. How I use every Claude feature (and how to set them up)

6. Design in Claude Code (not Canva or Figma)

7. Anthropic just dropped Claude Design

8. How I built a 2nd brain with AI

9. The 17 Claude skills behind 350k followers

10. Claude's features are wasting your time

11. Claude Code (for normal people)

12. Build your 1st AI agent in Claude Code (10 minutes)

13. Wait, your Claude doesn't have a harness?!

14. The Free AI Resource Vault (100+ prompts, Claude skills, workflows & guides)

15. Claude (actually) replaced my video editor

16. The 3 levels of Claude automation

17. Claude Code built my website (9 steps)

18. How to run Claude Code (for free)

19. NotebookLM + Claude Code

20. How to (actually) use Fable 5

21. Claude 5 vs ChatGPT 5.6

22. I made myself replaceable (and hired 3 people)

23. 'Record a skill' in Claude

24. Hermes Agent (Beginner → Advanced)

25. 100% humanised text (with AI)

26. Delete your CLAUDE.md

27. HTML > MD

That's it from me this week. See you again next week.

Stay curious, stay human and keep creating

— Charlie

"Show me."

The work existed. I just couldn’t see it with the wall of text that Claude would give me.

So I went and counted. Across 2,152 session transcripts and 15,109 prompts I had typed over 45 days, "show me" was my single most repeated phrase.

Three hundred repeats of the same two words isn’t a habit worth having.

Why it (actually) kept happening

The reason it kept happening was the format. Every answer arrived as Claudish, and Claudish is the one thing you can’t scan. I’m a visual guy, so I like to see things in front of me instead of reading chunks of text.

Someone at Anthropic wrote up the same problem in May, and that post has since done 14.6 million views and gone up on the official Claude Blog. It’s right, and I think most people took the wrong thing from it.

The read I kept seeing was "stop writing Markdown". People started converting their CLAUDE.md into HTML. That’s the one place it makes you actively worse off.

Here’s how I’m thinking about the split:

Markdown is what goes IN to the model.

HTML is what should come OUT of it, for you.

Drawn out, it’s this simple.

Put HTML in your config and you pay for it on every session you ever run.

Put Markdown on your screen and you get the wall of text you stop reading.

Last week was the other half of that audit, on the instructions themselves.

I walk through the whole setup in the video below, if you’d rather watch than read.

Markdown in, HTML out

Your CLAUDE.md is read before you type anything, on every session. Anthropic’s docs are blunt about what that costs:

"CLAUDE.md files are loaded into the context window at the start of every session, consuming tokens alongside your conversation."

Markdown is the leanest format Claude reads. HTML is the same content wrapped in tags that cost you tokens and buy nothing at all.

It’s called CLAUDE.md for a reason. There’s no CLAUDE.html.

When you ask Claude to build something, it reads all those Markdown files first. Then it has to tell you what it’s about to do.

That’s the moment it should switch, since the thing reading it is no longer a model with infinite patience but a human being with about four seconds of it.

Ask for a plan in the terminal and you get a wall of text to parse.

Ask for it as HTML and you get something you can scan.

For example, I’m doing lots of speaking over the next month, so I asked Claude to present me a visual plan. Now I can look at this at a glance and clearly understand my commitments.

I didn’t type any of those dates in.

Claude pulled them live out of my emails and my meeting transcripts, using Claude connectors. That’s the setting in Claude that lets it read from an app you already use, like Gmail or your calendar. I asked for the plan, and it went and got the data itself.

Type this in Claude Code

Read my calendar and my inbox for the next 30 days. Build a single HTML file: a Mon-to-Sun month grid, one cell per day, every commitment as a coloured chip. Work out the categories from what you find and put a legend across the top. Mark any double-booking in red, and flag any deadline that has no prep time blocked before it. Open it when it is done.

You need connectors switched on for that one, and nothing else. It reads the dates itself, so the only thing you are supplying is the shape you want them in.

It’s the same information. The difference is who is reading it.

Here’s another example I built this week.

That’s a real plan for a LinkedIn event I’m hosting this Wednesday.

The event description is written out in my voice, ready to paste. The cover is specced underneath it, sized and dated, with the file already sitting where I need it.

Everything was generated with Claude Code (even the image).

Type this in Claude Code

Build me a single HTML page for this event. The description written out ready to paste at the top, the cover image specced underneath it with exact pixel dimensions and the date on it, then a run of show as a timed table. Anything I still have to decide goes in a red block at the bottom. Open it when it is done.

You might have seen my infographics on LinkedIn. I’ve been designing them all programmatically with Claude Code (no designer, Canva or Figma).

I’m hosting a free LinkedIn live this Wednesday to show you how.

This time I’m going to be designing 4 infographics for you.

1. I’ll randomly select 4 people who repost this
2. I get two minutes to analyse each profile
3. You leave with a finished LinkedIn post

If you want me to build an infographic for you live.

Repost this, and I’ll randomly select four people on Wednesday.

What (most) people get wrong here

This is the cheapest upgrade you can make, and it only works once your output is visual.

Never ask for a version. Ask for variations, then pick.

When the output was terminal text this was useless. Comparing three walls of text is worse than reading one. When the output is HTML I can put three side by side, choose in about two seconds, and paste the choice back in.

You get three finished boards on one screen. You read them in about four seconds, click the one you want, and the build instruction are copied to your clipboard.

Type this in Claude Code

Build three versions of this graphic and put them on one HTML page side by side, same size, so I can compare them at a glance. Under each one add a pick button that copies the build instruction for that version to my clipboard. Do not describe them to me, just open the page.

For me personally, when I’m doing design work, I like to see multiple variations so I can pick and implement my taste. I’m not just outsourcing to AI here. I’m using AI to give me the variations, and I’m going to select the best one.

Once the graphic is done, Claude writes the caption and puts both on one page. The graphic on one side, the caption on the other, laid out the way the post will actually appear.

That saves the back and forth between Notion and the terminal, and it catches the things you only notice when they sit together.

Type this in Claude Code

Take the final caption and the graphic and build one HTML file laid out the way the post renders: image on the left at 4:5, caption on the right at the platform’s own line width. Under the caption, count the characters, the links and the hashtags, and flag anything outside the range my last thirty posts sat in. Open it.

I go for more than three variations if I need lots of options. For example, this is the plan for a YouTube video, and it ranked every title against live data before I looked at any of it.

I use Apify to scrape my YouTube thumbnails and some references.

I just pick the best one. The board is the shortlist, the taste is still mine.

Type this in Claude Code

Give me twenty title options for this. Rank them against real data, not your opinion — what has actually performed in this niche, how each one reads truncated at thumbnail size, and which promise the content can genuinely keep. Put them on one HTML page as a ranked table with the score and one line of reasoning beside each, top three called out. Open it.

Here’s another example from this week:

That board picked the cartoon at the top of this email. Three versions, drawn and laid out before I saw any of them.

The drawings came from Higgsfield, which Claude reaches on its own once you connect it.

Higgsfield is really easy.

1. Google "Higgsfield MCP"

2. Paste that into your Claude

3. Get a Higgsfield account

Now you have access to all the best models. Or add the ChatGPT API or the Gemini one instead. They both generate images. Claude does not.

It already knows what my drawings should look like, and that’s written down in a Markdown file too. I described the style once, and I’ve not described it since.

Show me → render it, then open it

The habit above all of this is simple. Stop typing more prompts and reading back through what happened, instead ask for the state of it instead.

My workflow is 4+ sessions open at once, Claude doing multiple things, and by the end of it I’m confused about where we left off.

The terminal isn’t very user friendly, and the desktop app is equally good, better in some ways even, I just find it slow.

So Claude’s been doing a lot of work in the background and all I have to do is ask it to show me where we’re currently at.

I’ve built that into a skill, so it fires whether I remember to ask or not. It builds the visual plan, renders it as a page and opens it in front of me. It’s free and it’s on GitHub: github.com/charlie947/show-me.

It renders the current state as one page and opens it for you.

Where you got to, what is built, what is left.

This has honestly changed my life, especially as someone who’s a visual person. I don’t like it when I’m just reading text. I need some visual stimulation.

Call me Gen-Z, call me whatever you want, but that’s just how it works for me and I stopped fighting it a while ago. If even a fraction of you have the same problem, you’ll find this useful.

Three things to install

Install 1. Make it show you instead of telling you

Everything above about HTML is a habit you have to remember. I got tired of remembering it, so it’s a skill now, and I’ve made it free.

Type this in Claude Code

/plugin marketplace add charlie947/show-me
/plugin install show-me@show-me

It clones the repo, checks it, and turns it on. From then on the two words do the work on their own.

Type this in Claude Code

show me

It clones the repo, checks it, and turns it on.

Then you say /show-me, and whatever you were working on arrives as a page instead of a paragraph.

Install 2. Make the diagrams stop looking like diagrams

The default when you ask any agent to draw something is a boxes-and-arrows chart that looks like every other one. This is the fix.

Type this in Claude Code

/plugin marketplace add cathrynlavery/diagram-design
/plugin install diagram-design@diagram-design

It gives you twenty seven editorial diagram types. Over sixteen thousand people have starred it.

Ask for a diagram after you install it and you will see the difference on the first one.

I drew that with the diagram skill above. I asked for a loop, it built the page, and it opened it in front of me.

Install 3. Make it lead with the answer

The last one is mine. It’s an adapted version of a skill called i-have-adhd by Ayoub Ghriss, and the credit sits in the repo.

Type this in Claude Code

/plugin marketplace add charlie947/answer-first
/plugin install answer-first@answer-first

It follows 11 rules. I've been running it for two weeks. Now I glance and know my next action.

That’s it running in a terminal session, and the bit I care about is right at the bottom.

What I love about it is where it puts things. I can be mid-session and what it gives me is Next, so I clearly know what I need to do without reading back up.

I just type /show-me or /answer-first depending on whether it’s a visual output or a conversation I’m having with Claude.

Bonus: Add this to your CLAUDE.md

HTML for deliverables: anything I VIEW or SHARE rather than paste into a platform (plans, code reviews, option grids, reports, dashboards, session handoffs) renders as a self-contained HTML file, inline CSS, no external requests. VIEW the rendered file before calling it done, and print its absolute path.

What stays Markdown: copy that gets pasted OUT to a platform (LinkedIn posts, the newsletter body, Notion page bodies) plus ALL core config, instruction and memory files. HTML there breaks the paste target or bloats token-priced context.

That’s effectively how I’m making AI feel less like AI. More like a co-pilot. Something that talks with me and speaks my language, rather than the AI jargon we’re all used to.

One ask before you go. If this saved you an hour, send it to a friend, or to whoever on your team is still scrolling terminal output.

That’s it from me this week. See you again next week.

Stay curious, stay human and keep creating

— Charlie

Not tidy it. Delete it. The CLAUDE.md, the hooks, the skills, 6+ months of instructions I'd written one bad session at a time.

The reason is that the man who built Claude Code told me to.

Anthropic deleted 80% of their own instructions

Boris Cherny, on stage at Y Combinator:

"We deleted 80% of the system prompt."

That's the prompt shipping inside the product, cut by four fifths for Opus 5 and Fable 5.

The model got better.

Then he turned it on you:

"For people that aren't building agentic products, but you're using Claude Code, every 6 months delete your Claude MD. Delete your skills. Delete your hooks. See what the model does and it might surprise you."

This was a revelation for me personally. I thought that all my old rules and instructions were still helping, when they had (actually) become a hindrance.

That's just one of three things I changed in my Claude Code workflow.

The other two are in the video below. Watch it first, then read the rest.

Before you scroll: if you find any of the following useful, someone on your team has the same problem and doesn't know it.

First, what these (things) actually are

I’ll explain this simply, because I got this wrong for a long time and the words get thrown around like everyone already knows.

Your CLAUDE.md is a system prompt. It’s the same idea as custom instructions in ChatGPT. A plain text file Claude reads before it reads your message at the start of every session.

Where I went wrong is what I thought it was for. I treated it as the place everything lives.

All my rules, memories, and instructions when actually it’s the front desk receptionist, not the filing cabinet.

There are five pieces and the thing that separates them is when each one runs:

CLAUDE.md → a standing brief

Claude reads it before you type a single word, every session, and you never trigger it.

Set it up with /init , then edit what it writes for you.

You get two, one personal and one per project.

That's a slice of mine of my personal Claude.md.

I have general rules that I want Claude to follow, such as entering plan mode first and delegating to sub-agents.

I also have a routing section that guides Claude to my other projects and .md files to get specific tasks done for me across my socials.

Haven't got one yet?

Genuinely, the best way to work is with Boris'.

That's (apparently) his whole file.

Mine is his file, 6+ months of corrections later.

Paste this:

Help me build my CLAUDE.md from scratch. Use Boris Cherny's CLAUDE.md as a starting template. Ask me about my business, voice, banned words, output defaults, and how I want you to work. Save the final file to ~/CLAUDE.md.

It interviews you. You answer. It writes the file.

When I tell Claude something is wrong, it writes it into its memory, not into my CLAUDE.md. And I run a skill at the end of a session that files those corrections into the right memory file for me.

It's a text file that grows every time you tell it something.

Now audit the one you already have

Mine wasn’t meticulously set up or planned. I built the whole thing by chucking context and ideas at Claude, all in natural language, over six months.

Which is why the audits are worth running. You end up with things you do not remember agreeing to.

/model

Do this before anything else. Things change. Some models want short instructions, some prefer verbose.

Some need examples, some get worse with them.

Then paste this. It reads Anthropic's live rules first, and goes through your file line by line.

It fetches Anthropic's live prompting guidance, finds where your chosen model excels, then updates your file against it.

It won't cut anything it can't justify with a quote.

Paste this:

My Claude instructions were written for older models. Audit them against Anthropic's current guidance and tell me what to delete.

1. FETCH THE LIVE RULES FIRST. Read Anthropic's prompting best practices page and the page for the exact model I am on. Use what is published today, never what you remember.

2. GO LINE BY LINE. One verdict per instruction: DELETE, KEEP or REWRITE, with the reason in plain English.

3. QUOTE YOUR SOURCE. Every DELETE quotes the sentence from Anthropic's guidance that justifies it. Cannot find one? The verdict is KEEP. Never invent a source.

4. FLAG EVERY VERIFY-TWICE RULE. "Always verify your work", "double-check before answering". The model already self-corrects. These make it do the work twice, and I pay for both.

5. FLAG "ONLY FLAG THE BIG ISSUES". Taken literally, so I get told less than I wanted. Replace with: report everything, I do the filtering.

6. FLAG ANY "DON'T OVERTHINK" RULE. Rules against thinking make internal tags leak into the answer. Cut them.

7. FLAG ROLE PADDING. "You are an expert with 20 years of experience" was once good advice. Now optional. Keep a role only where it genuinely changes the output.

8. FLAG STALE EXAMPLES. Examples written for an old model teach old habits. One current example beats ten stale ones.

9. NEVER TOUCH A TRUTH RULE. "Only claim what you verified" is not a severity filter, it stops made-up facts. Rules like that stay, whatever else goes.

10. TELL ME WHAT IS MISSING. The new models need five instructions the old ones did not: keep answers short, cap document length, say how to update me while you work, hold the task scope, limit the helpers you spawn.

THEN GIVE ME:
- A table, one row per instruction: my line, the verdict, the why, and Anthropic's line.
- The exact wording to paste in for each of the five missing instructions.
- The honest count. "307 lines in. 4 deletes, 4 rewrites, 5 additions."
- What you could not check, marked NOT RUN. Never a clean bill you did not earn.

AND BE HONEST:
- If most of my file should go, say so plainly. The scaffolding I am proudest of is the likeliest casualty.

My instructions are below. Go.

It goes and reads Anthropic's live docs before it judges a single line of yours.

Then it hands back a plan. This is mine, run this morning against the 282 lines I have today:

Nothing gets deleted the moment you paste it. It backs up your .claude folder first, and both routes hand you a plan to approve.

Read it properly before you hit yes.

Skills → a workflow you saved

A skill is a prompt you saved and gave a name to. You write the workflow once, then you type one slash command instead of retyping the whole thing, in any conversation, in any folder.

I had 195 of them. They cost nothing sitting there, because only the one you call actually loads. Which is why the problem is never how many you have. It's how many are doing the same job.

The prompt that sorts this out is in the audit run further down.

Here's the prompt I'd run first.

Audit my skills folder for overlap and dead weight. Do not delete anything.

1. List every skill with its name and one line on what it does.
2. Group them by the JOB they do, not by their name. Skills doing the same job in different words are the same skill.
3. For each group of two or more, tell me which one to keep and why, and exactly what the others have that it is missing.
4. Flag every skill I have never actually used.
5. Flag every skill that only works because of a file or folder that no longer exists.

Then give me ONE merged version of the biggest group, written out in full, ready to save. Tell me what I lose by merging.

It goes through your skills, finds the ones doing the same job, and tells you which to keep. It deletes nothing:

Out of the 195 skills on my machine, I’d only ever invoked 65. That’s why this audit is essential.

Hooks → an automatic trigger

This one tripped me up for a long time.

If a rule in your CLAUDE.md is a request, then a hook is a rule that fires whether Claude fancies it or not.

The simplest one you could build:

When I say the word “LinkedIn”, write me a LinkedIn post.

That's a hook.

I never tell it to. I just say the word/s.

This is where hooks and skills meet. The word is the hook. What it goes and gets is the skill. You cannot see that pairing anywhere in Claude, which is why it took me so long to work out I already had loads of them.

Start by finding out what you already repeat. Paste this:

Help me work out what I should turn into a hook.

Look back over our recent conversations and find:

1. Things I ask for over and over, in roughly the same words.
2. Times I had to tell you something you should already have known.
3. Times you did something I had to correct more than once.

For each one, tell me:
- the word or phrase I actually say when I want it
- what you should go and read, or do, the moment I say it
- whether that word is unique enough that it won't fire by accident

Give me your top three, ranked, with the best one first. Do not build anything yet.

It read 15,109 prompts I had typed over 45 days. My most repeated phrase was "show me", 424 times, and nothing was firing on it.

Then just tell Claude to turn the top one into a hook (no special command or config needed). You describe it, it builds it, and it shows you before it saves anything.

Worth adding this to your CLAUDE.md while you're there:

If I ask you for the same thing three times in a session, tell me, and suggest it as a hook.

And once you have a few, audit them. This lists everything, including the ones that came with plugins:

Audit every hook I have set up. Do not change or delete anything.

1. List them all. For each one: what triggers it, when it fires, and what it actually does, in plain English. No file paths.

2. Tell me which ones fire on EVERY message. Those are the expensive ones, so I want them justified or gone.

3. Flag any trigger word that is too common. If a word shows up in normal conversation, the hook will fire when I did not mean it.

4. Flag any two hooks that could fire at the same time, and tell me what happens when they do.

5. Flag any hook pointing at a file or folder that no longer exists.

6. Tell me which ones I have probably forgotten I set up.

Then give me a table: keep, fix, or remove, with one line of reasoning each. Do not touch anything until I say so.

What (most) people get wrong

We keep adding more and more instructions instead of pruning things back.

That hurts the long term capability of your Claude.

The more instructions you write, the more bloated your CLAUDE.md becomes, and the more confused the model gets over time.

Anthropic publish a ceiling for this and you probably didn’t know that it:

"Size: target under 200 lines per CLAUDE.md file. Longer files consume more context and reduce adherence."

Mine was at 305. That was a mistake, because these aren’t settings.

They're suggestions competing with each other:

"Claude treats them as context, not enforced configuration."

And a few lines later:

"if two rules contradict each other, Claude may pick one arbitrarily."

Now imagine ninety-six rules, written across six months, by one person correcting whatever annoyed him that day. Some of them contradicted each other and I never noticed.

There's a second effect underneath, and it's why a long file makes your prompt *worse* rather than just pricier.

Models read the beginning and the end of a long context carefully, and they skim the middle.

The paper that named it is called Lost in the Middle.

So a bloated CLAUDE.md costs you three things at once:

1. Tokens. Every line is paid for on every single session, whether it's relevant or not.

2. Accuracy. Rules that contradict each other get resolved arbitrarily, and you never see which one won.

3. Attention. Your file goes in first, so every line you add pushes what you actually asked for further from the edge the model reads carefully.

The whole thing, in the order it fires

Here is what actually happens when I ask for something. Four steps, and I only do the first one.

1. I type. My global CLAUDE.md is already loaded. It knows who I am, how I write, and where everything lives. That is the only file that costs me on every session.

2. A word in my message fires a hook. Say carousel, say thumbnail, say Substack.

3. The hook goes and gets the project file for that job. The infographic one is 860 lines, and it never enters a newsletter session.

4. That project file points at the skill that does the work, and the skill runs.

I never start in the right folder. I open Claude on my desktop like everyone else, because I cannot be bothered to navigate first.

That costs me tokens. But the hook does the walking, so I get the right rules anyway.

That is the point. Your main file should not hold the knowledge. It should point at it, and something else should do the pointing for you.

But like anything it needs maintenance and top-ups over time.

Nothing ever seems streamlined and efficiently running forever without maintenance. So no, I'm not deleting mine. I'm maintaining it.

That's it from me this week. See you again next week.

Stay curious and stay human

— Charlie

I’m writing this newsletter (without AI). Something I rarely do. I’ve never been a “great” copywriter (I hated English in school).

AI has helped me share my thoughts online and it’s probably the reason why I’m here today.

But I've managed to get it down to a way that generates 100% human generated text with AI. The whole workflow is at the bottom of this issue.

Substack’s AI detection

Substack just launched the AI detection feature in collaboration with Pangram and from what I’ve seen, the initial reaction has been negative.

Boy, I write on LinkedIn, and it’s funny the CEO of Substack posted this. LinkedIn just released a “seems like AI slop” feature.

This genuinely worries me because, as someone who uses AI every day in almost all aspects of his content business, I fear that people might use this in an attempt to just be anti-AI generally.

Because I used AI to create a post, does that make it slop? I wrote a whole article about this, but in my opinion, slop is low-effort AI content. Content without a point of view. Content without any personality.

So in essence, I do agree with Chris here.

Claudefishing is a fun new word. I’ll definitely be using that 🤣

Anyway, the negative reaction. What seems to be the problem here is that AI detectors never truly do work. For example, back in April 2023, a popular AI detection tool, Zero GPT, thought that the US constitution was mostly written with AI.

I don’t know what they were using back in 1787, but this definitely isn’t AI.

If anything, I think there's gonna be a resurgence in em dashes.

I think it is a fancy way of writing that actually shows that you know how to write. Obviously, they've had a rough few years thanks to AI using them all the time, because AI was trained on that dataset.

AI was trained on good writing.

There are false positives across a lot of AI detectors for non-native speakers, and autistic writers too, for repeated phrases. But also for people who aren't good copywriters.

And of course it's going to be the same with false negatives. With Pangram, apparently there's one in seventy false negatives.

40.5% of LinkedIn posts over 250 words are being fully AI-generated.

I'm definitely culpable for some of them. But hey, LinkedIn really is turning into a bit of an AI slop place.

Especially when I look at the comments section.

I'm just like, gosh, people, do you not have original thought? Or at least find a way to use AI effectively to write good comments, instead of just sounding like an absolute robot.

How I (actually) write my newsletter

I genuinely thought I had a sick process.

I’ve been refining my newsletter process with Claude Code, baking in all my previous editions so the writing is based on my data and the specific writing style that I’ve developed over the past two years of being on this platform.

I then use the gstack plugin, and it enables the /browse feature to write the entire edition for me based on the context I provide it with. Recently, I’ve even recorded YouTube videos and then got Claude to write the entire newsletter, screenshots and all, based on the video input.

For example, this edition follow that process:

But then that still comes back 100% AI 😅

Because yes, technically, it is AI-assisted. I recorded the video, and then Claude Code did the rest for me. But does that make the newsletter any less human?

I recorded the video. Everything is my original thought and demonstration.

All right, so I’m around halfway through writing the article, and currently Pangram is saying this is 100% human written (which is correct).

So what about if we deploy ways to see if we can evade the detection by writing the next section in three different ways using AI.

So I’m going to test four ways with the exact same article from last week (which was flagged as 100% AI-assisted) and see if we can evade detection.

Deep research in Claude

Okay, so first I conducted a deep research report in Claude to identify how it can beat the Pangram Substack AI detection.

It gave me back a comprehensive report, and I just simply asked Claude to turn that into a Claude skill.

So we actually ended up on two skills here:

AI actually explains it better than I do, so take that AI detection.

You can’t detect screenshots of AI outputs :D

If you want to try both of the skills, you can access them here.

Humaniser prompt

So I used this viral prompt from to see if it would truly humanise the text.

# FOLLOW THIS WRITING STYLE:

• SHOULD use clear, simple language.
• SHOULD be spartan and informative.
• SHOULD use short, impactful sentences.
• SHOULD use active voice; avoid passive voice.
• SHOULD focus on practical, actionable insights.
• SHOULD use bullet point lists in social media posts.
• SHOULD use data and examples to support claims when possible.
• SHOULD use "you" and "your" to directly address the reader.
• AVOID using em dashes (-) anywhere in your response. Use only commas, periods, or other standard punctuation. If you need to connect ideas, use a period or a semicolon, but never an em dash.
• AVOID constructions like " ...not just this, but also this".
• AVOID metaphors and clichés.
• AVOID generalizations.
• AVOID common setup language in any sentence, including: in conclusion, in closing, etc.
• AVOID output warnings or notes, just the output requested.
• AVOID unnecessary adjectives and adverbs.
• AVOID staccato stop start sentences.
• AVOID rhetorical questions.
• AVOID hashtags.
• AVOID semicolons.
• AVOID markdown.
• AVOID asterisks.
• AVOID these words:

"can, may, just, that, very, really, literally, actually, certainly, probably, basically, could, maybe, delve, embark, enlightening, esteemed, shed light, craft, crafting, imagine, realm, game-changer, unlock, discover, skyrocket, abyss, not alone, in a world where, revolutionize, disruptive, utilize, utilizing, dive deep, tapestry, illuminate, unveil, pivotal, intricate, elucidate, hence, furthermore, realm, however, harness, exciting, groundbreaking, cutting-edge, remarkable, it, remains to be seen, glimpse into, navigating, landscape, stark, testament, in summary, in conclusion, moreover, boost, skyrocketing, opened up, powerful, inquiries, ever-evolving"

# IMPORTANT: Review your response and ensure no em dashes!

Here’s the analysis:

That worked surprisingly better than I expected. We’ll take 31% human here.

Humaniser skill

This one is on GitHub, and it follows the AI writing tales from Wikipedia. I honestly expected it to perform worse.

Training AI on my human written text

I expected this to work a lot better. Maybe the sample size was too small here. Because 99% of my content online is AI-assisted text, I don’t have any pure human-written text at this point, besides this article.

What (actually) works

I've been testing this workflow for the past few hours.

So I wanted to build one that actually works and actually writes like you.

So every step on how to humanise your writing with AI and get 100% on the detection with the Pangram detector on Substack is there.

So I've created this repo on GitHub. It's free, and it got me to 100% human.

You can install it with this command:

git clone https://github.com/charlie947/voiceprint.git

Because sometimes when we prompt AI, we just give it a top line overview and then we expect it to make up all of the connecting dots. Whereas in this instance, Claude is going to keep drilling down until it gets enough context and enough spoken words to convert into a newsletter.

Effectively here we're not using AI as a replacement. We're using it as a thinking partner to help flesh out our thoughts, the own human content, to then turn it into a newsletter which sounds like us and is human written.

Now we're gonna use Pangram to see whether it's AI. I'm quite nervous about this because I haven't tested this.

Voila, it's 100% human.

This is definitely a way that I'm going to be creating my newsletters in the future.

But ultimately, the point I want to make for writers online is that if you're a writer, this isn't for you. If you enjoy the craft of writing, then keep doing that.

If you're someone like me who prefers to communicate visually, whether that's on camera or with designs, then this is the perfect solution.

So, as a conclusion here, I don’t really think it bloody matters if your Substack is AI-generated, AI-assisted, or human-written.

What matters is that you show up online, providing it comes from a place of lived experience and an actual first-person point of view.

If AI removes the barrier for you to share your thoughts online (like it did for me). Then great, I’m happy for you because it can really change your life.

What truly does matter is the reader’s experience.

So I’d like to ask you these questions directly:

If you missed it, I recorded LinkedIn Live showing you exactly how I make my infographics (because I’m a visual guy, not a written text kind of one).

This is the same system that runs my entire content operations across multiple platforms. You can try copying the output, but as I always like to say: “AI can copy the output. It can’t clone the chef”.

So if you’d like to be involved when it officially launches, just sign up here.

That’s it from me this week. See you again next week.

Stay curious and stay human

— Charlie

This is a collaborative issue. Sabahudin Murtic gets you set up from scratch. I show how I run Hermes day to day. Jérémy Grandillon takes it advanced. Three levels, one newsletter.

Everything here is how we actually run Hermes, not theory. If you are new here, welcome. If you have been reading a while, thank you. Your subscription keeps this going.

In this issue:

• Set up your own Hermes from scratch.

• See exactly how I run mine every day.

• Push it to autopilot with scheduled routines.

I recorded how the system actually works in a quick video (under 10 minutes).

Watch it first, then read the rest.

So what actually is Hermes?

It’s a free, open-source AI assistant that lives on your own computer, or a separate Mac Mini.

It remembers you, and it does real work on its own. You text it like a contact in your phone, and it gets on with the job while you do something else.

You might already know OpenClaw. It went viral this year, half a million people run it, and it put the whole idea of an agent you own on the map.

I tried it. For me it didn’t stick. It was buggy, things broke mid-task, and I spent more time fixing the agent than getting work out of it.

Hermes is the same idea. The difference is it just ran.

• It stays stable, backed by the lab behind the Hermes models.

• One login for any model, cheap for volume or local for privacy.

• Do a task once and it writes itself a skill, so it never has to be told twice.

So it stuck. Now it replies in my WhatsApp, runs real jobs on my Mac behind an approval gate, and checks my team in on Notion, every day.

OpenClaw is the one most people have heard of. Hermes is the one I reach for.

Now let’s build one, beginner to advanced.

Hermes (Beginner)

by Sabahudin Murtic from Unstuck

Everyone in my feed lost their mind over Hermes this year.

200,000+ stars on GitHub. The number one agent on OpenRouter. NVIDIA even wrote a whole post about it.

So before Charlie and Jeremy take you into the deep end, let me be the guy who explains it from zero.

What it actually is

Hermes is a free, open-source AI assistant you install on your own computer or a cheap $5 server. It lives there and stays on.

Three things make it different from regular Claude or ChatGPT:

• It remembers you → Every session. You stop re-explaining your business every single time.

• It improves itself → When it figures out a task, it writes its own little “skill” and reuses it later.

• It runs on its own → You text it on Telegram or WhatsApp, and it does the work while you sleep.

That’s it.

An AI you own, that remembers you and actually does things.

Why people go crazy for it

Because it’s yours. Free forever. Your data stays on your machine, not on someone’s server.

And you talk to it like a contact in your phone, not another browser tab you forget to close.

For a solopreneur, that is a cheap, private, always-on assistant that actually knows your business.

What it costs (the honest math)

The software is free. It runs on a $5-a-month server. You only pay for the AI thinking it uses, usually a few dollars, not another subscription trap.

Some people run their whole setup for under $20 a month.

Now let me be real

Hermes is not one-click. It is an autonomous agent. That means it can run commands on your computer on its own.

Powerful? Yes. But you do not hand a stranger the keys to your house.

So you set it up safe: keep the approval settings on, run it on a cheap throwaway server, and lock it so only you can message it.

And my usual line still holds: do not chase the tool before you know your problem. You cannot automate what you have not mapped. Hermes will not fix a messy business. It will just run your mess faster.

Who it is for right now

If you can open a terminal and paste one command, you are fine.

If that sentence scared you a little, wait. Or get someone to set it up with you first.

The simplest way in

One line in terminal:

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

In your terminal should appear this:

Your first setup (5 quick settings)

Once it is running, here is the basic setup. Five things, then you are done.

1. Pick its brain. Choose which AI runs it.

hermes setup --portal

2. Turn on memory. So it remembers you, then tell it what your business does.

hermes config set memory.memory_enabled true

3. Put it in a safe box. Sandbox it and keep approvals on.

hermes config set terminal.backend docker

4. Lock one channel to you. Link Telegram and let only your own ID in.

hermes gateway setup

5. Give it one job. Tell it a recurring task in plain words, like: every weekday at 9am, summarise my inbox and send it to me.

Stuck? Run:

hermes doctor

It tells you what is wrong.

Which model I run, and why

Hermes lets you plug in any AI as its brain, and even give it a cheaper second model for the grunt work. This is where the money quietly leaks. Here is my setup.

I run two models, not one:

→ Opus for planning. The smart, pricey one thinks and makes the plan. Planning is a small slice of the work, so paying premium here is cheap, and a good plan saves you from expensive mistakes later.

→ Sonnet for execution. The fast, cheaper one does the actual work: reading, drafting, running the steps. Execution is where most of your tokens burn, so this is where the real savings live.

The rule is simple. Expensive model where the thinking happens. Cheap model where the volume happens.

Want to go further? Send the boring background jobs (titles, summaries, cleanup) to the cheapest model you can find. They do not need a genius.

Why does it work? You pay per token. The plan is small and high value. The doing is huge and repetitive. Match the model to the job and you get almost the same output for a fraction of the cost.

That is it. A smart brain to decide, cheap hands to do, and a tiny model for the chores.

Another setup worth stealing: cheap or private

Claude is my pick, but it is not the only game. Depending on what you care about, here are the two other setups people run.

→ Cheap: DeepSeek V4.

Almost frontier quality for a fraction of Claude’s price, with a discount on repeated context. Want it smart and cheap? This is the one.

→ Private: an open model on your own machine.

Run Qwen or Llama locally with Ollama. No token bill at all, and your data never leaves your computer. The catch: you need a decent GPU, and it is a little weaker on the hard stuff.

Not sure and do not want to juggle keys?

Nous, the makers, recommend their Portal: 300+ models (Claude, GPT, Gemini, DeepSeek, Qwen) on one subscription so you can try any of them.

The point: the model is a dial, not a default. Premium for quality, budget for cheap, local for privacy. Same Hermes, very different bill.

Want the deep version? Nous has a full model guide here: hermes-agent.nousresearch.com/docs/user-guide/configuring-models.

After you set up everything correctly, type “hermes” to start.

Connect Telegram and you are texting your own AI and you are ready to go.

That is the beginner map from my side. Charlie takes it from here.

Hermes (Intermediate)

by Charlie Hills

Once it is set up, Hermes stops being a novelty and becomes staff.

First, I gave Hermes its own number. It needs a separate WhatsApp, not your personal one, so I bought a cheap second number from Hushed (about $6 a month), connected it to WhatsApp, and added it to my seven team group chats.

Second, I gave mine a personality, a name (Hermione), and one hard rule she can never break: she is never me. An agent that thinks it's you is dangerous.

All of it lives in one plain-text file called SOUL.md. Think of it as a CLAUDE.md for who your agent is, not how it codes. It holds her voice, her style, and the lines she must never cross, all in plain English.

It sits at ~/.hermes/SOUL.md, and Hermes writes you a starter version the first time it runs. You don't create it, you just open it and make it yours. Change the file, change how she behaves.

Now, the brain. I don't run a frontier model on her. Mine runs on GLM and DeepSeek, two of the cheapest, fastest models going.

I went with Nous Research's own subscription (around $20 to $100 a month) rather than juggling individual API keys, so I can switch models whenever I want.

Most of what she does all day doesn't need a big brain. Reading, logging, chasing, pulling numbers. Cheap and fast handles it, so the bill stays tiny and she never leaves me waiting.

Finally, permissions. Hermes only replies to the numbers I've allowed, nobody else.

Without that, anyone who landed in a group chat could set it running, and it's doing real work on my Mac. So I locked it to my own number and kept the approval gate on.

1. Claude Code in my pocket

I wired Hermes into Claude Code on my Mac. Now I can text it from WhatsApp and it runs jobs on my machine. Everything happens behind an approval gate.

But she never runs off on her own. Before anything executes on my Mac, Hermes shows me the exact command and waits. Nothing runs until I approve it from my phone.

One approval per job. That gate is what makes handing a WhatsApp agent real work safe instead of reckless.

I did not hand her the whole machine.

She can run exactly two jobs on my Mac through Claude Code. One is research, read-only, for pulling numbers and comparisons. The other is team check-ins, where she reads the team's tasks in Notion and logs their updates. Nothing else. She cannot touch my files, my keys, or her own config.

2. Content research on tap

We drop links from across social into one channel.

She doesn't just summarise it. She scrapes each post with Apify, works out how far it beat that account's own baseline, and has Gemini watch the video.

Then she files the winner in Notion with the multiple attached. A 20x outlier is a very different signal from a 2x, and now I can see which is which at a glance.

Two connectors, one API key run this for me.

Open the connector directory, search for each one, hit the plus, and sign in through the browser. They show up in Claude Code automatically.

Apify → scrapes any social media.

Notion → stores everything (our 2nd brain).

Gemini → for multimodal (video) analysis. You need an API key for this.

Grab a free one at aistudio.google.com/apikey

Once done, drop it into a .env file at the root of your project:

GOOGLE_AI_API_KEY=your_key_here

That’s the whole setup. Text this:

Here is a reel: [link]. Break down why it works and add it to our content research.

3. Team check-ins

A teammate taps in and Hermes logs it. It reads their tasks in Notion, our second brain, and reminds them what is still open. The chasing I used to do, now done without me.

Everything lives in our “Flamingo Fam Hub”.

Meetings, SOPs, tasks, schedules - basically everything Claude and Hermes needs to run operations live in here.

You don’t need a fancy “second brain” like Obsidian.

Claude + Notion MCP is the only combo you need.

Hermes (Advanced)

by Jérémy Grandillon

Alright, now that Sabahudin and Charlie got you started, it’s time to push it to the next level!

We’ll use two main concepts here:

• Sabahudin explained that your Hermes agent has its own computer (either one you physically own or a dedicated server you own/rent)

• Charlie told you that you should treat Hermes as a team member.

A special teammate

Hermes uses its computer through the terminal, which allows basically to do anything you would do using the human interface and even more.

One thing that makes Hermes special is that it never sleeps, and it can create crons.
If you’re not familiar with cron, it’s basically routines.

So now you may be seeing me coming and start thinking of all those things that you and your human team never prioritise due to lack of time and resources.

Yes, you guessed it.

Hermes is your personal assistant by day and works on scheduled tasks at night.

Set up crons

Honestly, you’ve done the heavy lifting already with Charlie and Sabahudin.

Now all you have to do is chat with your agent using your favourite channel and just prompt it to do something regularly.

Example of a message I sent my agent to set up a routine.

I personally have a bunch of routines running at night (great way to optimise your subscription usage I recommend it!), so I have what I need in the morning.

Some are content-oriented. Others are for admin stuff, and others are just for my personal life.

Examples of some of my content routines.

Self-improvement

We know that now AI is as good as its context.

And this obviously also applied to your Hermes agent, but nobody wants to feed context manually. That’s annoying.

This is where routines come in: give your agent scheduled learning sessions while you sleep.

It can scan your content like I’m doing, but it can also go through your internal databases like Charlie showed you with Notion.

Now you have it: your agent is always up to date, going through your source of information and updating its own knowledge files.

And as it stays on its own computer, security breaches are very low risk.

Giving Hermes the right tools

Alright, this is where it becomes juicy.

Giving your agent connectors and MCP is already great but I find it a little limited sometimes.

Having my agents go through my tools using APIs, CLI, MCP is cool, but sometimes I want something different. I want my agent to act like a real team member with its own credentials.

In reality, you want a mix of that. And it will also depend on how you like to work, your taste, your goals, etc.

Quick example of the tools I gave my Hermes agent:

• Notion: limited to the pages I choose, through a Notion app

• Accounting tool: API in read-only, so my agent can remind me when something is late in payment

• Email account: call me old school, but I don’t want everything in Telegram. I use AgentMail, so not only can Hermes send and receive emails, but it can also create accounts in tools

• Apify: through API, Charlie already shared this gem, basically allowing your agent to scrape anything on the web (note that “actors” in Apify vocabulary are small cloud programs that do a specific automation job).

• GWS: I do need Hermes to read my emails sometimes to find info or files in that mess

Note: this isn’t an official Google CLI, but it’s made by Google’s team and has been tested loads of time. Just sent the github link to your Hermes agent and follow its lead!

And I keep adding and removing stuff as my usage and habits evolve. Talking daily with my agent.

It uses a bunch of tools that I’m not aware of, and that’s okay. I don’t expect my assistant to wait for me to solve problems.

It’s actually the point: if it can do it without you, it will just do it.

Wait, isn’t that what Claude Slack tag does?

Not really.

The idea is similar in some ways, but there are some fundamental limits/differences:

• Hermes is model agnostic (or almost), so you don’t have to have Claude

• As Sabahudin told you earlier, Hermes is an open-source project. So it’s basically free to use, and you own it fully

• Tag is only available to teams or enterprise plans (for now at least), so not everybody has that

• ⁠From what I’ve seen (not 100% sure), tag is less flexible. You can’t give the tools you want, just the connectors available in the store

And what about sales? Aren’t you the GTM guy?

Yep, that’s me.

And you’re right. I didn’t mention too many GTM plays I do with my Hermes agent, and the reason is simple: I’m not using it that much for those use cases.

I do sometimes send my Hermes agent a Telegram message with a LinkedIn profile so it can come up with something that my prospects have said in their content so I can shine in a call.

I also ask it to update things in my systems.

But as I am the GTM guy, I already have a bunch of dedicated agents and workflows that are doing this for me, so I keep my Hermes agent as my personal assistant.

And that would be my final tip: create your Hermes agent, and use it daily.

Ask yourself if your assistant could handle the task you’re facing right now.

This is the only way you’ll see where in your life, business, and existing systems your Hermes fits and shines.

Run the whole thing yourself

One agent that remembers you and does the work is where this starts. Wiring it into the whole operation, the research, the check-ins, the jobs that used to need me in the room, is the system I run my business on now.

That is the LinkedIn AI OS, and I just opened it up today.

It is in Alpha, so I am taking people in small batches.

If you want the exact system behind my LinkedIn, apply below.

Apply to the LinkedIn AI OS Alpha.

One ask: only apply if you have opened Claude Code before. This alpha is for people helping me refine it before the wider launch.

Want to see it run first? I am hosting a free live session on exactly how I use it.

Save your seat.

Stay curious, stay human, and build the one teammate you actually own.

— Charlie, Sabahudin & Jeremy

P.S. New here? Subscribe free and my Resource Vault (100+ prompts, Claude skills and workflows) lands in your inbox the moment you join.

Ways we can work together

Become a Paid Subscriber: Get full access to every past issue of MarTech AI and support future ones.

Claude Content Engine: The Claude engine behind 350k+ followers (across socials) and 100 million impressions a year.

The exact system that now builds every AI infographic on my profile, GIFs and all. Two years in the making, and it is finally out. Check it out here.

For two years I have done one job entirely by hand.

Last week's issue showed you the system that turns one idea into content for every channel.

The last piece still on me was the scheduling. Every post gets re-cut for Threads, X and Facebook and dropped into Buffer, one platform at a time.

So I recorded myself doing it once. Claude watched the whole thing and turned it into a skill that now runs it for me.

No prompt or instructions typed out. I showed it, and it built the skill.

That is Anthropic's new Record a skill. If you have never made one, a skill is just a job you teach Claude once and run again on command.

Normally you teach it by writing the steps out. This lets you show them instead, and it lives in Cowork on the Claude desktop app (you need a paid plan, Pro, Max or Team).

In this issue:

• How to record your first skill, start to finish

• The exact words to narrate while you record, ready to copy

• How to turn it into a task that runs itself every morning

If you're new here, welcome. If you've been reading for a while, thank you. Your subscription is what makes issues like this possible.

I recorded the whole thing in a quick video. Watch it first, then read the rest.

Record your first skill

You would never hand a new teammate a hundred numbered steps and walk off. You would sit with them for ten minutes and show them.

Claude learns the same way now. Nail the recording and you nail the skill, so here is how to get it right.

Step 1: Open Cowork and start recording.

Open the Claude desktop app and select Cowork. Click the plus icon and choose Record a skill.

Claude asks permission to record your screen and watch what you do. Read the warning first: do not type passwords or secrets, or show private conversations, while it is rolling.

Then hit Start Recording.

Step 2: Do the job once, for real.

Pick something you actually repeat. I chose content repurposing, the one job I have never automated.

I opened the post in my Notion calendar, grabbed the caption and the image, then went to Buffer to schedule it.

Then the fiddly part. On Threads a long caption has to be split into a proper thread. X needs a tighter version. Facebook keeps the longer form. I did all of it by hand, exactly as I do every week.

Step 3: Narrate as you go, and stay on task.

Say what you are doing and why, out loud. Claude reads the screen and listens, and the why is what it turns into the skill.

The one habit that makes or breaks it: do not get sidetracked. Claude captures every single click. Wander into an unrelated tab and it assumes that is part of the job too, then repeats it every time you run the skill.

Step 4: Stop, and let Claude propose the skill.

Hit stop. Claude takes the recording and gets to work.

It read my demo back to me first:

“I can see from your recording that you're building a repeatable content-repurposing workflow... The narration makes clear you want this captured as a reusable skill. Let me set that up.”

Then it split the build across three sub-agents.

1. One drafts the skill file.

2. One writes a per-platform reference (the Threads-splits-but-Facebook-does-not rules).

3. One packages it and hands it back.

Out came a skill it named repurpose-to-buffer, with the exact steps I showed it. One nice touch: it did not just record my mouse clicks to replay them. It worked out the proper tool for each step, reading Notion through a connector and driving the browser only where it had to.

Step 5: Save it, then test it before you trust it.

Press Save Skill. Then open a fresh chat and run it with a slash command, /repurpose-to-buffer (or whatever Claude named yours).

Testing now is your one chance to refine it before it fires for real.

It runs through Claude in Chrome, which can do almost anything a person can in a browser. In the desktop app, turn on Control Chrome and grant the permissions as they pop up.

Without it, Claude cannot act on your Chrome, and nothing happens.

Then watch it go. Claude opened Buffer, pasted the caption, split the thread, ticked the right channels, and queued the post. The job I have done by hand for two years, running itself.

Say this while you record

The recording is only as good as your narration. Replace anything in [BRACKETS] with your own words, and talk through it like this:

Say this:

I'm doing [TASK] so that [OUTCOME].
First I [ACTION], because [REASON].
Now I [NEXT ACTION], and the reason is [WHY].
This next bit is the tricky part: [THE RULE THAT CHANGES BY CASE OR PLATFORM]. That is the whole job. Anything I clicked that was not part of it, ignore.

Say the why for every step. That is the part Claude turns into the skill.

Make it run every morning

Testing it once is nice. A system is better.

So I gave it one line.

Paste this:

Let's make this a scheduled task. Ask me when it should run, then set it up.

Claude asked how and when I wanted it to run.

Then it built the scheduled task, every step written out, set to run on weekdays at 8am.

One honest caveat. Because this drives your real browser, your Mac has to be awake and signed into Buffer for it to run. This is not the cloud kind that works while your laptop is shut.

If the machine is asleep when it fires, it will not half-publish. It sends you the prepped post and stops.

Show it, don't tell it

The skill you build this way is not a prompt. It is a recording of you doing the job right, once.

Type instructions and you get Claude's best guess at your process. Show it and you get your process.

The catch is the same one as always. It copies exactly what you do, the clean parts and the sloppy ones. So do the boring job properly one last time, then hand it over for good.

Run the whole thing yourself

Recording one skill is how you retire one manual job. Stringing them together is the system I run my content on: the research, the writing, the design, all of it.

That is the LinkedIn AI OS, and I just opened it up today.

It is in Alpha, so I am taking people in small batches.

If you want the exact system behind my LinkedIn, apply below.

Apply to the LinkedIn AI OS Alpha.

One ask: only apply if you have opened Claude Code before. This alpha is for people helping me refine it before the wider launch.

Want to see it run first? I am hosting a free live session on exactly how I use it.

Save your seat.

Stay curious, stay human, and show it, don't tell it.

— Charlie

P.S. New here? Subscribe free and my Resource Vault (100+ prompts, Claude skills and workflows) lands in your inbox the moment you join.

Ways we can work together

Become a Paid Subscriber: Get full access to every past issue of MarTech AI and support future ones.

Claude Content Engine: The Claude engine behind 350k+ followers (across socials) and 100 million impressions a year.

For two and a half years, I have poured everything into LinkedIn.

Don’t get me wrong, I love the platform and love making content. It grew my audience past anything I expected, but if I want to build something bigger than me, I have to get out of my own way. The founder doing every task is the ceiling. I was the ceiling.

This year, I’m doing something about it:

1. I made my first key hires (shoutout to team Flamingo).

2. Handed over 80% of my content work to them and Claude.

I spent months feeding Claude my tastes, my standards, the way I judge whether a post is worth publishing. Everything I carry in my head now lives in the system.

For 2+ years, “everything” lived in a mess of notes on my phone.

Now the idea leaves my head and goes straight to an agent on WhatsApp.

Because nobody really struggles to execute anymore (AI made that part easy). What you lack now is ideas and a system that turns them into posts worth reading.

This is the most comprehensive system I have built.

Today I am showing you exactly how it runs.

In this issue:

• How the system decides what is worth writing.

• The exact process that scores an idea against your own data.

• The structure that turns guessing into a systematic approach.

If you’re new here, welcome. If you’ve been reading for a while, thank you. Your subscription is what makes issues like this possible.

I recorded how the system actually works in a quick video (under five minutes).

Watch it first, then read the rest.

This is what I (actually) do

It runs like a team now.

Every channel I publish on, LinkedIn, Instagram, YouTube, this newsletter, goes through one operator, gets scored by one brain, and is remembered in one place.

I think of the tools as a department each. One runs operations, one researches, one writes, one designs, one keeps the memory. I set the direction and sign all of it off.

Here’s exactly how it (actually) runs, and how you build a version of it this week.

Three tools carry it.

1. Notion is the memory.

2. Hermes is the operator.

3. Claude Code is the executor.

I set up Hermes, Nous Research’s open-source agent, with a single command:

1. I gave it a spare phone number (a cheap VoIP line from Hushed)

2. Registered WhatsApp on it, and linked Hermes as a device.

3. It runs quietly on a Mac mini, off my main machine.

If you never heard of Hermes (not the luxury designer), then don’t worry.

Next week’s newsletter is going to show you exactly how to set it up.

I made a Content Research WhatsApp group and added my team.

Then I connected my Claude to my Notion:

1. Open Settings and go to Customize, then Connectors.

2. Find Notion in the list and click Connect.

3. Authorise and pick the workspace.

After that, Claude can read and push everything to my Notion.

Stage 1: Research comes in

My team and I drop stuff into the WhatsApp group.

We tag Hermes and it pulls each post (through Apify), it works out how far it beat that account’s own baseline, and files the winner into Notion.

We just dump ideas and inspiration in there without losing them.

You don’t need the WhatsApp agent to start. You need your own posts in one place first. So export them. On LinkedIn I use AuthoredUp to pull my posts and their real numbers to a CSV.

On Instagram I use Apify. Then I hand that file to Claude and let it build the dashboard.

Paste this:

I have connected you to my Notion. Attached is my post export [ATTACH IT: your AuthoredUp CSV from LinkedIn, or your Apify pull from Instagram]. Build me a Best Posts database with the post, topic, hook type and its numbers, a Voice page for how I write, and a Content Calendar. Fill Best Posts from my export, then tell me what you built.

Now your best posts live somewhere Claude can open. Score your ideas against them.

Stage 2: The winners become a scored queue

The winners land in my Ideation Hub in Notion, ranked.

When I ask what to post, an agent reads my own best posts and generates new angles that already work for me, each ranked by tier.

What comes back is not a blank-page draft. It is modelled on a piece of mine that already earned its reach. I pick from a shortlist stacked in my favour.

Stage 3: Claude creates it

I hand the top idea to Claude Code. It already carries my content bible, my voice file and my story, so it never starts cold.

I created these because I was (honestly) tired of re-explaining myself to Claude every time, so at the start of each session, these are automatically loaded up.

It writes the post, then a separate agent scores that draft against a rubric built from my own winners and rewrites until it clears 95.

I never see a first draft.

Paste this:

Using my Best Posts and my Voice page in Notion, take the top idea and write today’s post in my voice. Model it on my best-performing post on a similar angle. Use only facts I give you, invent nothing. Then score your own draft out of 100 against my winners and rewrite until it clears 90. Show me only the final version and its score.

Example output:

I analysed my last 254 LinkedIn posts.

Three hooks drive almost all of my reach.

The median post gets 21 reactions and 4 comments.

My best posts do over 100x that.

The difference is rarely the topic. It's the first line.

Here are the three hooks that win for me:

1. Free resource

- "Anthropic won't send you this list."
- "My entire Claude skills library is now free."
- These earn more reposts than anything else I publish.

Give away something useful. People share what makes them look smart.

2. Pain point aimed at you

- "You are burning through your Claude credits by lunch."
- "You're paying for Claude and using 10% of its brain."
- Name the reader's problem in the first six words.

The sharper the pain, the faster they stop scrolling.

3. Stat or curiosity gap

- "95% of Claude users stop at Claude Chat."
- "24 tools inside Claude Code run my entire business."
- Open with a number that begs a question.

Curiosity buys you the second line. The second line buys the rest.

Everything else I post lands near the median.

These three pull it far above.

Save this before you write your next hook.

Repost ♻️ to help someone in your network.

P.S. Which of the three do you open with most?

Stage 4: The design builds itself

The same idea goes to my infographic generator. It’s an agent team that writes the copy, builds the graphic in code, and runs it through a brand gate before I ever look.

Every post ships with a graphic that looks like mine.

Example output:

Can you believe that Claude generated this? I can. I’ve spent the last four months building out the system. If you want to be the first to try it, sign up here.

Stage 5: One idea becomes every channel

A LinkedIn winner does not stop at LinkedIn.

It gets logged in a repurpose queue and re-cut for every platform.

LinkedIn

Instagram Carousel

Instagram Reel

What it hands me is not a word-for-word script. It is a package I can riff off the cuff, a spoken hook, an on-screen hook, and the talking points.

YouTube runs its own pipeline on top of that, each newsletter becomes a video.

All of this is (almost) done automatically within Claude Code.

Stage 6: The team stages it, I publish

It all lands in my Notion Content Calendar.

Claude writes to it, Hermes reads from it, and I approve from it.

Then the team takes over.

One person owns LinkedIn, the infographics and the carousels. Another owns Instagram, the B-roll and the edits. And I am slowly handing this newsletter over too.

None of them are me.

None of them need to be.

• Because I built the system.

• Because we all work from the same shared folders.

• Because I spent the months putting my taste, my voice and my judgement into code, into files the whole team runs from.

What used to live only in my head now lives in the system. So the work is completely outsourceable to people who are not me, and it still comes out looking like mine.

I am the final gate, and only the final gate.

I read it, I feel it, I approve it.

That last yes is the one thing I keep.

Stage 7: The system sharpens itself

Every post comes back with a number. Once a week I feed those numbers in.

An eval recomputes my median, tags every post that beat it by 5x, flags the flops, and rewrites the one file that tells the system what is working for me right now.

So it never coasts on last month’s winners. It chases this week’s.

• The scoring gets better.

• The angles track what is actually working.

• The writer snaps back to my voice every time it drifts.

I just keep publishing. The machine keeps tuning itself to the results.

A prompt is only as good as the day you wrote it. This gets better every week I use it.

So, should AI replace us? Yes. I think it should. My output grew so fast that I hired three people this year just to run the system with me.

The AI did not take those jobs. It created them.

That is the version of replaceable I am chasing.

— Charlie

P.S. If there is a section you want me to cover next, tell me in the poll below or reply to this email. I will expand on it next time, because you decide what this newsletter covers next.

Ways we can work together

Become a Paid Subscriber : Get full access to every past issue of MarTech AI and support future ones.

Claude Content Engine : The Claude engine behind 350k+ followers (across socials) and 100 million impressions a year.

The Claude guy.
The king of orange.
The bloke who started a full-blown romance with Claude in his own comment section.

I've earned all of it. For six months, this newsletter has been Claude, Claude, Claude.

Then GPT-5.6 landed, and it was impossible to avoid.

The benchmarks were genuinely stunning. OpenAI hyped it for a solid week, and I sat and watched the launch livestream, which I haven't bothered to do for a launch since GPT-5.

So it earned a real test. Not a benchmark score I have to take on faith (I can't see behind those charts either), but the actual jobs you'd throw at it on a normal Tuesday.

So I ran it against Claude Fable 5 on the only jobs that matter (the ones I actually do).

If you're new here, welcome. If you've been reading for a while, thank you. Your subscription is what makes issues like this possible.

So I gave both tools the same five jobs. I gave the exact same prompt to each (no cherry-picking or second attempts I quietly deleted).

I cranked the effort dials to the max on both and stopped worrying about tokens.

GPT-5.6 Sol

Quick map if you're not deep in this. GPT-5.6 comes in three tiers: Sol at the top, Terra in the middle, and Luna, the fast, cheap one. I ran Sol on Extra High.

You pick the model and the effort dial in the picker at the bottom of the chat.

Claude Fable 5

Claude's lineup has more names. Fable 5 is the new flagship, the one built for your toughest challenges, with Opus 4.8 just beneath it, Sonnet 5 for everyday work, and Haiku 4.5 for fast and cheap. I ran Fable 5 on Max.

I bought three 20x Claude Max plans this week and they reset on Thursday, so this was the week for it.

I thought f*ck it, let’s burn these tokens, baby.

Test 1: A website

First one's a laugh. I exported my LinkedIn profile to a PDF, dragged it into both tools, and gave them the same brief.

Build me the best portfolio website you can from my LinkedIn profile. Go crazy with the animations. Make it look expensive.

Fable 5’s version

The nicer overall experience, the one I'd actually live in.

GPT 5.6's version

Slicker animations, I actually preferred them here.

Call it a tie. ChatGPT's animations edged it, Claude's overall feel edged it back.

Honest bit though: I one-shot both of these in a chat, which is exactly what you shouldn't do for anything real. One-shot a website and it'll always smell a bit off. You build it properly, with a system → Claude Code is terrible at design.

One point back to ChatGPT though. It has ChatGPT Sites now, which builds the site and hosts it on a live link you can share. Claude hands you the file and leaves the hosting to you.

Test 2: GTA 6

Then I got greedy.

A game either runs or it doesn't, so I skipped the sensible brief and asked for the most ridiculous thing I could type with a straight face. And I told it not to fob me off with the old top-down rubbish.

/goal build me GTA 6. Proper 3D, not the 1997 top-down version. An open world with real streets I can drive a car around, proper lighting, buildings, the lot. Use WebGL, and make it as realistic as you possibly can. It has to run the second I open it.

Fable 5’s version

Prettier world, nicer lighting, the one you'd screenshot.

GPT 5.6’s version

Less pretty, but more complete: real missions, cops that chased me, felt alive.

Fable for the showreel, GPT-5.6 for the one you'd actually play.

This is the one to run with /goal in Claude Code or Codex. You set the bar (make it actually drivable, make it 3D) and it keeps rebuilding and checking its own work until it clears that bar, then stops on its own.

Test 3: Minecraft world

Then the fun one. A Minecraft-style world I could actually fly through, a proper flight sim, not a still.

Build a breathtaking, scroll-stopping flight simulation through an
endless Minecraft-style voxel world, as ONE self-contained HTML file.

Hard requirements:
- Stage is exactly 1080x960. Design everything for that square frame.
- Single file. No external assets, no libraries, no CDNs, no fonts, no
  network requests. Everything hand-coded (raw WebGL or canvas).
- Plays the instant the file opens. No clicks, no loading screen.
- The camera flies forward forever over procedurally generated terrain
  that never repeats and never ends. No player input needed — the
  flight path is cinematic: smooth banking turns, gentle climbs and
  dives, low passes over ridgelines.
- The world must feel VAST: long draw distance, wide vistas, terrain
  fading into atmospheric haze at the horizon.
- Terrain variety on camera within the first 30 seconds: rolling plains,
  forests with trees, a river or lake, snow-capped mountain peaks.
  Blocks are visibly blocks — crisp voxel edges, grass tops on dirt,
  stone above the treeline, snow on summits.
- Lighting sells it: a low golden-hour sun, blocks shaded per face,
  soft shadows across valleys, drifting blocky clouds above, subtle
  god-rays when the camera turns toward the sun.
- Smooth motion — hold a steady high frame rate over that draw
  distance. Stutter fails the brief.

You are judged on the sense of scale first, then beauty of the light,
terrain variety, voxel charm, and smoothness of the flight.

Fable 5’s version

The nicer colour and brightness, softer and easier on the eye.

GPT 5.6’s version

Way richer. Looks like the real thing, smoother transitions, the lot.

Genuinely torn here. A richer world or prettier light? Call it a tie.

Test 4: A pitch deck

Build me a 5-slide pitch deck for an app that texts you one genuine compliment every morning. Make the design go hard. No corporate stock-photo energy.

Fable 5’s version

Nailed it. The kind of deck I'd actually present.

GPT 5.6’s version

I said no corporate stock-photo energy. It handed me corporate stock-photo energy.

No contest. Claude took it.

Test 5: An infographic

This is the one where I stopped being the judge.

I ran both through my own system, the one that scores a graphic against my brand rules before I even look at it.

Create an infographic comparing Claude Code and Codex.

Fable 5’s version

This is my actual system, the one I run in Claude Code. First-gen and honestly not bad. I'd make a few edits, but a solid generation.

GPT 5.6’s version

Loved this one. It used one of my existing templates, the styling was way better, and the richness of information was near perfect.

Genuinely a tie. And that's the point: when both are this good, it comes down to the harness. Which is exactly what my LinkedIn AI OS is.

That system is one piece of something bigger I've been building: my LinkedIn AI OS. Brief in, brand-locked graphic out, the whole pipeline I run my content on.

It's in Alpha. Rough in places, and the cheapest it'll ever be.

Join the LinkedIn AI OS Alpha and use code ALPHA for 50% off while it's this early.

One ask, though: don't join if you've never touched Claude Code. This alpha is for competent users helping me refine it before the wider launch.

Use both, don't pick

Everything above tied. So the only thing left to compete on is the cost, and that's where this gets interesting. There's a Codex plugin that drops OpenAI's Codex right inside Claude Code, and it's what I've been running all week.

You plan with Fable 5 or Opus, hand the grinding to Codex, then let Fable review what Codex wrote. Best model planning, cheapest one on the labour, best model checking it.

And the price is the whole point. Fable 5 runs about $10 in, $50 out per million tokens. Push the work to Codex instead and you drop to a fraction of that, Terra at $2.50/$15, Luna at $1/$6. Same output, a slice of the bill.

So which one won?

GPT-5.6 wins the lab, or so the launch says. OpenAI's own chart puts Sol at 88.8% on Terminal-Bench, about five points clear of the Fable 5 you can actually buy at 84.3%. But squint at it: the Claude model right on Sol's tail at 88.0% is Mythos 5, and you can't get your hands on it. A benchmark makes a lovely headline. My five builds were the real test, and they tied.

Build the harness once, then pour in whichever model's winning this month.

Stay curious, stay human, and combine, don't pick.

— Charlie

P.S. The poll at the below decides next week. Vote, then hit reply and tell me the one job you wish your AI did on its own. I read every one.

Ways we can work together

Become a Paid Subscriber: Get full access to every past issue of MarTech AI and support future ones.

You don’t have long with it though. It’s included in your Claude plan until 7 July, then it moves to pay-as-you-go credits.

While you have it, don’t run it like a smaller, older Claude with a bigger context window. It’s more powerful, but that’s not the point. The point is loops.

One of them ran on this very edition.

It built a rubric from my 110 past editions, caught this draft at 51/100, and rewrote it until it passed. You’re reading the pass.

In this issue:

• The exact method to run your first loop today

• The eval loop I use to make Claude check its own work

• The rubric that scored this very edition, built on my own data

If you’re new here, welcome. If you’ve been reading for a while, thank you. Your subscription is what makes issues like this possible.

What is a “loop”?

Every loop has the same four beats.

1. Goal. What “done” looks like, written down before anything runs.

2. Act. The model takes one step towards the goal.

3. Check. Something compares the result against the goal.

4. Repeat. If the check fails, go again. If it passes, stop.

A prompt is one turn. A loop runs until the check passes. The check is the beat everyone skips, and I’ll keep coming back to it.

How to prompt Fable 5

Fable 5 isn’t a faster model. It’s a different one, and it needs different prompting.

Anthropic published the full guide. The most useful bits are paste-able. Here are the ones I actually use:

Give it the reason, not just the task.

Fable works better when it knows the purpose. Ten extra words, far more output.

I’m working on [LARGER TASK] for [WHO IT’S FOR]. They need [WHAT THE OUTPUT ENABLES]. With that in mind: [YOUR REQUEST].

Tell it when to act.

It can think for minutes, which is great on hard problems, but on simple jobs tell it when to stop planning.

When you have enough information to act, act. If you’re weighing a choice, give a recommendation, not a survey of every option.

Make it check its own progress.

Long runs used to fake status updates, so make it check every claim against the real result.

Before reporting progress, check each claim against an actual result. Only report work you can point to evidence for. If it isn’t verified, say so.

Set boundaries.

Fable sometimes takes initiative you never asked for.

When I’m thinking out loud rather than asking for a change, give me your assessment and stop. Don’t make changes until I ask.

Keep it running on long jobs.

Deep in a long session it can stop early, or panic about running out of context.

You’re working autonomously and have ample context. For reversible steps, proceed without asking. Don’t stop or suggest a new session.

Give it a memory.

It works best writing lessons to a file and reading them back. A markdown file is enough.

Keep a notes file. Store one lesson per entry with a one-line summary of what worked and what didn’t. Update existing notes instead of duplicating.

The rest are settings, not prompts. The one that matters is effort. It replaces model switching. Instead of dropping to a smaller model for cheap work, you turn the dial down.

This is Anthropic’s own benchmark. Fable 5 on medium beats Opus 4.8 on every setting, including max, and costs less doing it.

High is the default and the sweet spot.

The jump to xhigh buys five more points on the hardest problems. Max costs half as much again for barely two. Leave it on high, and push to xhigh when the job is genuinely hard.

Max is for the run you’d rather overpay for than redo.

How I actually build a loop

You don’t start with the loop. You start with the eval, a written test for the work so the loop can mark its own output instead of asking you.

Writing this very edition, I built one.

I write my newsletter in Claude Code, and a QA agent reviews every draft before I do.

First I needed the data.

I had Agent Browser, Vercel’s automation browser for AI agents, log into Substack and export all 110 of my past editions with their open rates and engagement.

Then Claude compared the winners to the flops and turned the difference into a scoring rubric.

Now a fresh agent scores every draft against it, cold. Fable plans, Sonnet executes, and the judge stays separate so nothing marks its own homework.

Under 95, it goes back with notes. This edition went from 51 to 95 before I read a word.

And it compounds. Every failure I catch by hand, I add to the rubric, and the loop never misses it again.

Build your own. Paste this:

Here are my three best [posts / emails / designs]: [PASTE THEM]

1. Study them and write a scoring rubric out of 100. Base the categories on what makes these good: [hook, structure, voice, clarity]
2. Draft a new one on [TOPIC]
3. Score the draft against the rubric as a strict reviewer
4. Under 95: rewrite and score again. Repeat until it passes
5. Show me only the final version and its score

That’s the whole loop in one chat. Build your rubric and yours does the same.

Your data is the moat

Every AI has read the same internet, so everyone’s output sounds the same. It hasn’t read your engagement stats, your best work, your flops.

Your own data is the moat. Feed it into the loop and the work is yours, not everyone’s.

The check is the job now

Start with one small loop this week. Claude Code has two commands for it:

1. /goal sets the finish line in one testable sentence

2. /loop keeps Claude working until that goal is true

The first loop feels like more effort than doing the task yourself. The tenth time, you’ve forgotten the task existed.

Writing the loop is the easy part. Writing the check is the job.

A prompt fails in front of you and you fix it on the spot. A loop fails while you’re asleep, handing you 50 finished pieces of work that are all subtly wrong.

The check you wrote is the only thing between you and that pile. So the human doesn’t leave the room. The human moves, and the job becomes:

• Define “done” before anything runs

• Read the checks, not the drafts

• Keep your hand on the publish button

Slop is a failure of judgement, not tooling. Loops make that truer. They’ll produce more of whatever you tell them to accept.

Run this system yourself

I’ve just opened the alpha for the exact system in this issue.

The loops, the rubrics, the QA agent, set up to research, write and design your content (infographics and animated GIFs included).

I’m only letting 20 people in. Use code ALPHA at checkout for 50% off. First come, first served, and it’s the cheapest it will ever be.

Claim your spot.

Stay curious, stay human, and write the loop, not the prompt.

— Charlie

Ways we can work together

Become a Paid Subscriber: Get full access to every past issue of MarTech AI and support future ones.

Claude Content Engine: The Claude engine behind 350k+ followers (across socials) and 100 million impressions a year.

It interviewed me on my business and my content.

Every answer cited the exact episode it came from.

Then I handed it my LinkedIn analytics, my Substack stats, and my content numbers.

It found the gap in 30 seconds. I reach 500k+ people a week and I own almost none of them. My followers are LinkedIn's, not mine.

You can do this two ways. Clone an expert’s thinking and it mentors you. Clone an expert’s workflow and it works for you.

By the end of this newsletter, you will have built both.

If you are new here, welcome. If you have been reading for a while, thank you. Your subscription is what makes issues like this possible.

MarTech AI is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

You already follow the right people

You follow the right people. Entrepreneurs, the operators, the marketers you actually trust.

You save the podcasts. You watch the videos. You screenshot the good bits.

None of it changes what you do on Monday morning.

Two things break here.

The first is advice.

You ask ChatGPT or Claude what it thinks and it averages the whole internet back at you. It has never met your business. It does not know your numbers. So it guesses, and the guess sounds confident, and you act on nothing.

The second is the workflow.

A creator shows you the exact process on a Tuesday. You nod. You save the video. You never run it once.

The fix for both is the same. Stop consuming the expert. Clone them.

The Expert, Cloned Two Ways

Every expert you follow is two things at once.

They are a way of thinking. And they are a set of workflows.

You can clone both, and you use two tools to do it.

Mode 1: Mentor

Their thinking becomes advice. Grounded in their real words, cited back to the source, pointed at your business.

Mode 2: Skill

Their workflow becomes a tool. A saved process you run on demand, built from a single video.

What you need:

• A free NotebookLM account

• Claude Code installed on your laptop

• Your trusted advisors’ podcasts or videos

• Your own business data (analytics, revenue, numbers)

Half one builds the mentor. Half two builds the skill. Start with the mentor.

Mode 1: Mentor

Step one → Load your advisors in NotebookLM

NotebookLM is Google’s research tool. It is free. You give it sources, and it answers your questions using only those sources.

That last part is the whole point. A normal chatbot invents an answer when it does not know. NotebookLM tells you the answer is not in your material. It will not make up advice in your advisor’s name.

So you give it the real material from the people you trust.

The setup:

1. Go to notebooklm.google.com and create a new notebook.

   

2. Name it after your advisor or your theme. Mine is “Diary of a CEO”.

   

3. Add sources. Paste YouTube links to the episodes, or upload transcripts and PDFs. I loaded 300 Diary of a CEO episodes.

   

4. Wait for it to process. Each source gets read and indexed.

5. Ask your first question. “Based on these episodes, what would you tell me about pricing a new offer?”

   

The answer comes back grounded. Every claim carries a little numbered citation.

Click it and you land on the exact moment in the exact episode where it was said.

Three hundred Diary of a CEO episodes is not one mentor. It is 300 interviews with some of the best operators alive. James Clear is in there. Naval Ravikant is in there. So your notebook is not Steven Bartlett. It is a room full of advisors, and you pick which one answers.

So I ask each one the question they are best at.

James Clear, on habits.

I asked how to hold a daily writing habit when I travel and my routine falls apart.

It came back with his two-minute rule. Shrink the habit until it is too small to skip. Write one sentence, not one article. Then it added his point about identity. You do not write to publish, you write to become someone who writes. Both quoted from his Diary of a CEO episode and linked, so I could go and hear him say it.

Naval Ravikant, on leverage.

I asked how a one-person business grows without hiring a team.

It pulled his three forms of leverage. Labour, capital, and products that cost nothing to copy, which he sums up as code and media. A team is the old leverage. A post that works while you sleep is the new one. It cited the exact passage. I run a content business, so that changed what I do with my mornings.

Steven Bartlett, on what to cut.

I asked what I should stop doing in my content.

That answer became the audit in step three, so I will show you the whole thing there.

No invention. No generic filler. Their actual thinking, on my actual question.

Use this when: you want to think through a decision in a specific person’s frameworks, not a blended average of everyone.

The limitation: NotebookLM lives in your browser and it is slow once a notebook gets big. It is brilliant for thinking. It is not built to sit inside your daily workflow. That is what step two fixes.

Step two → Connect your notebook to Claude Code

This is the step that makes the whole thing work, and it is the one most people skip.

Right now your advisors live in NotebookLM and your work lives on your laptop. They cannot talk to each other. This step connects them.

Claude Code runs skills. A skill is a saved capability you switch on once. The one you want is called notebooklm-py, an open-source skill built by Teng Lin. It lets Claude Code ask your notebook questions directly, so you stop copying answers back and forth by hand.

The setup:

1. Install it. In your terminal, run

   uv tool install “notebooklm-py[browser]

   No uv on your machine? Use

   pipx install “notebooklm-py[browser]

2. Log in with

   notebooklm login

   This opens your browser to sign in with Google once.

3. Add it to Claude Code with

   notebooklm skill install

   That drops the skill straight into Claude Code.

4. Point it at the notebook from step one with

   notebooklm use YOUR_NOTEBOOK_ID.

Set this up once. Every question from here runs through it. Now ask Claude Code to put a question to Bartlett, and the answer lands in your terminal, citation and all.

Use this when: you want your advisors available inside the place you actually do the work, not in a separate browser tab.

The limitation: the skill drives your browser session, so keep NotebookLM open and logged in while you work.

Step three → Run the audit

Your advisors are loaded and connected. Now you point them at your real numbers.

Claude Code can read the files on your laptop. You give it two things. Your advisors, and your business.

The setup:

1. Drop your real data into a folder. Your LinkedIn analytics export, your newsletter stats, your content numbers.

2. Run one audit prompt.

Here is the prompt I used.

You have access to my Steven Bartlett "Diary of a CEO" advisor notebook and my business data
in this folder. Act as my advisor.

1. Read my content performance numbers.
2. Find the single biggest gap between what I do and what the data says works.
3. Back your answer with the specific Bartlett episodes that apply.
4. Give me one change to make this week. Be specific. Use my numbers.

It read everything and came back with one gap.

I reach a lot of people. Last week my posts hit 513,858 impressions and reached 202,081 people. Followers up 2,274.

Almost none of that is mine. My email list, the only audience I actually own, got 319 clicks in the same week. That is 0.16% of the people I reached.

Then it pulled the backing from the notebook. Daniel Priestley, on Diary of a CEO, says capitalism is about ownership, and renting your audience from a platform is a losing position. Bartlett makes the same point about owned data over rented reach. Both episodes were cited, so I could check them.

The call was simple. Stop pouring everything into reach you rent. Turn some of it into an audience you own.

So here is the one change. My top post last week reached 145,751 people at a 6.9% engagement rate. If 2% of one post like that joins my email list, that is around 2,900 owned subscribers from a single post. Nearly ten times what my whole site got all week.

I am taking the advice. There is an email capture on this very issue.

What this gives you:

1. Advice grounded in someone you trust, not a model’s guess.
2. Advice pointed at your real numbers, not a generic best practice.
3. One specific action, not a list of twenty.

I wrote about asking AI better questions back in:

This is the version with a brain and your data behind it.

The limitation: the audit is only as honest as your data. Feed it vanity metrics and it advises on vanity. Give it the numbers that actually matter to you.

Bonus → Convene a board instead

That whole build is an afternoon. Loading episodes, connecting tools, pointing it at your data. Worth it once. Too much when you want an answer today.

So I built the fast version.

It’s a Claude skill called board-of-advisers. You name the mentors you admire. Claude becomes them. You ask one question, and each one answers in their own thinking. Then a fifth voice, The Chairman, reads every answer and hands you one verdict with a first move.

No NotebookLM or transcripts need.

The setup:

1. Download the skill (below)

2. Open Customise skills in Claude.

3. Upload the SKILL.md file.

4. Type “convene my board”.

Then name your board and ask. I put one question to mine.

Four advisers answered.

Hormozi said fix the offer first.
Welsh said the newsletter is fragmented, so double it.
Dan Koe said pick the one product that lets you leave the work.
Naval said only the SaaS scales without your time.

Then The Chairman weighed all four.

The chairman’s verdict

Where they agreed: my audience is the asset. Where they clashed: Hormozi wants revenue now, Naval wants leverage later, both right on different clocks. The first move it gave me was simple. List my revenue and my hours per product, side by side. The focus becomes obvious.

One Claude agrees with you. Four advisers tell you what you missed.

Use this when: you want the board in 30 seconds, not the full build above.

The limitation: the board is fast, but it does not read your data files the way the audit does. For a gut-check on a decision, it is perfect. For a numbers audit, run Mode 1.

Mode 2: Skill

Step four → Turn any YouTube video into a Claude skill

The mentor advises you. A skill does the work for you.

You already met skills in step two. There you installed one. Here you build your own.

A skill is a saved instruction file. Claude reads it and follows the steps every time, so a workflow you learned once becomes a tool you keep forever.

The source can be any tutorial video. A creator demonstrates a process on camera. You turn that process into a skill, and now you can run it without watching the video again.

The setup:

1. Find a video that demonstrates a workflow you want to keep.

2. Pull the transcript. (YouTube shows it under the video, or use a free transcript tool.)

   

3. Paste it into Claude Code with one instruction.

Here is the transcript of a video that demonstrates a workflow.

1. Identify the exact step-by-step process being taught.
2. Write it as a reusable SKILL.md I can run again on new inputs.
3. Keep my decisions as inputs. Automate the mechanical steps.

Claude reads the transcript, strips out the chat and the intro, and writes a clean skill file of the workflow itself.

I tested this on one of my own videos. It pulled the workflow I demonstrate on screen and turned it into a skill that runs the same process on a fresh topic. Readers can build that skill straight from the video and keep my method, instead of re-watching it every time they need it.

That is the part I like. You stop being a viewer of the workflow and start being an owner of it.

Use this when: a creator shows you a process you want to run more than once.

The limitation: the skill copies the steps, not the judgement. It will run the process. It will not know when the process is the wrong call. You still make that decision.

Bonus #2 → Clone yourself, then publish

Mode 1 clones the mentor. Mode 2 clones the workflow. There is a third clone, and it is the one that turns all of this into reach. You clone your own voice.

Look at what you are holding now. A notebook full of advisor thinking, and an audit full of insight about your own business. That is a month of content sitting in a folder. The audit told you to grow the audience you own. Content is how you do it. The only thing between the folder and the feed is the writing, and the writing is the slow part.

That is the job Stanley does. It learns your voice from your existing posts, your vocabulary, your structure, your rhythm. Then you hand it raw material and it drafts the post in your voice, not generic AI filler. You edit 20 percent instead of rewriting 80.

So the bonus move is simple. Take what your board said, or what the audit found, paste it into Stanley, and let it draft the LinkedIn post that summarises it. The thinking is yours. The voice is yours. Stanley does the typing.

Paste the notebook → draft in your voice → publish

I ran it on this exact issue. One paste, one draft, a post about cloning your advisors, written the way I write. Try Stanley free for 14 days

What this does not solve

Garbage sources give garbage advice. The advisor is only as good as the material you load. Load three random clips and you get three random opinions.

A skill is a snapshot. When a creator updates their process, your skill does not. You refresh it.

And the setup is an afternoon, not five minutes. Loading 300 episodes took time. Connecting the tools took a few goes.

You still own every editorial call. The advisor advises. The skill executes. You decide.

Pick one creator this week. Build the mentor and build the skill from the same person. Watch what they tell you to fix, then build the tool that fixes it.

Stay curious, stay human, and stop shipping the default.

— Charlie

P.S. Who would you clone first? Reply and tell me the one advisor you would put in the notebook.

Ways we can work together

Become a Paid Subscriber: Get full access to every past issue of MarTech AI and support future ones.

Claude Content Engine: The Claude engine behind 350k+ followers (across socials) and 100 million impressions a year.

By 5:21pm Eastern, Fable 5 and Mythos 5 were gone. Not throttled. Switched off, for every customer, everywhere.

By the end of this issue you’ll have his playbook, plus my own verified setup, so you can run capable AI on your own machine, for free, with no company and no government able to flip the switch.

No YouTube video from me this week. I'm writing this from Cannes, coffee in hand and a beautiful view of the bay in front of me. I’ll be back in July.

If you’re new here, welcome. If you’ve been reading for a while, thank you. Your subscription is what makes issues like this possible.

MarTech AI is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

You don’t own the model

This was the first time the United States has ever issued an export-control directive for access to an AI model. The order barred foreign nationals from using Fable 5 and Mythos 5, including Anthropic’s own staff who aren’t US citizens. To comply, Anthropic had to pull both models for everyone.

I’m not here to argue whether the order was right. Plenty of people are doing that already. The lesson sits underneath the politics, and it applies to you whichever model you pay for.

You don’t own the model. You rent access to it.

Greg Isenberg put it best, and this whole issue is built on the point he made on LinkedIn this week. His framing: don’t build your entire workflow on something that can disappear with one letter. Own part of your stack.

He’s seen this film before. He once ran a startup whose biggest traffic source was organic Facebook. The algorithm changed overnight and he lost 99% of it. Anyone who built on rented reach got the same lesson. Email lists exist because people learned it the hard way.

You can already see it happening, and not just with hobbyists. PewDiePie, one of the biggest creators on the planet at 100 million plus subscribers, just went fully local.

He built a 10-GPU rig at home, ran a stack of open models against each other to pick the best answer, and released his own self-hosted AI workspace called Odysseus (https://github.com/pewdiepie-archdaemon/odysseus) with his own model in the works.

He owns the whole thing rather than renting it. Odysseus is an agent harness, the same idea I broke down in.

When the biggest channel on earth builds its own, you know where this is going.

So here’s how to setup your first local LLM this weekend.

Local models are insurance. You keep using frontier models for the heavy lifting. You keep a local fallback for the day access vanishes, or you need privacy, or the wifi drops on a train and you still want to work.

I call the setup the Local Stack. Eight steps, in order.

1. Pick your runtime (Ollama or LM Studio).
2. Match the model to your hardware.
3. Pick the right model for the job.
4. Shrink it with quantisation.
5. Point at your local model.
6. Mind the context window.
7. Give the model tools.
8. Fine-tune later.

What you need

• A laptop (a Mac with 16GB or more of memory is ideal)

• Ollama or LM Studio, both free

• One model downloaded to start (a small one)

• Twenty minutes

The first five steps get you running. The last three make you good at it.

If you want the secret sauce behind how I generate GIF infographics like this, apply here.

Step one → Pick your runtime

A runtime is the program that runs a model on your own computer. It’s the engine. No runtime = no local AI.

You have two free options. Ollama is the one I use. Download it from ollama.com, install it like any app, and it quietly runs in the background.

LM Studio is the other, with a friendlier window if you’d rather click than type.

Once Ollama is installed it serves models at a local address on your machine, localhost:11434. You don’t need to understand that line to use it. Just know it’s the door everything else knocks on later.

To download your first model, you run one line:

ollama pull [model]

That pulls the model onto your hard drive. From then on it runs offline.

Tip: Start with one small model before you download a 40GB monster you’ll never use. You can always pull more.

Step two → Match the model to your hardware

A model’s size is measured in billions of parameters. You’ll see numbers like 7B, 27B, 70B. Bigger is smarter, and bigger needs more memory.

Here’s the rough rule:

• A 7B model runs on almost any laptop

• A 27B to 32B model wants 32GB or more of memory

• A 70B model needs serious hardware, like a maxed-out Mac Studio

If a model is too big for your memory, it spills over and runs three to five times slower. So match the model to the machine you actually own, not the one you wish you had.

Use this step when: you’re about to download a model and you’re not sure your laptop can handle it. Check the size first.

Step three → Pick the right model for the job

Different open models are good at different things. These four cover almost everything a marketer or operator needs:

• Qwen 3 is the best all-rounder for most tasks.

• DeepSeek is the one for reasoning and coding.

• Gemma 3 is tiny enough to run on a phone.

• Llama has the biggest community.

You’re not locked in. Pull two or three, test the same prompt across them, keep the one that answers best for your work.

Tip: The “best” model on a benchmark is rarely the best model for you. The right model is the one that runs smoothly on your machine and gives you usable answers. A fast 7B you actually use beats a 70B you never load.

Step four → Shrink it with quantisation

This is the one concept that changes everything, and almost no one outside the field knows it.

Quantisation shrinks a model so it runs on weaker hardware, with barely any drop in quality. When you browse models you’ll see versions labelled Q4 or Q5. Those are the shrunk ones. Q4 keeps roughly 95% of the quality at a fraction of the memory.

This is how a model that “needs” a server ends up running fine on your laptop. Look for the Q4 version by default. Learn this word and the whole thing opens up.

Step five → Point Claude Code at your local model

This is the payoff, and it’s the part I verified myself across all those versions.

Claude Code is the terminal tool a lot of us now build with. By default it talks to Anthropic’s cloud. You can point it at a model running on your own machine instead.

You don't need to touch a config file. With Ollama installed, one line does the whole thing:

ollama launch claude

A menu pops up. Arrow-key to your local model, press Enter. Behind the scenes Ollama points Claude Code at your machine instead of the cloud, so your requests never leave your laptop. Now the tool runs on a local model. Free. Offline. Yours.

Tip: you’ll want a recent Ollama (v0.15 or later) for the launch command. Update first if ollama launch isn’t recognised.

Want more than Ollama’s one-liner? There’s a popular open-source proxy called free-claude-code (https://github.com/Alishahryar1/free-claude-code), 35k stars and counting, that points Claude Code at 17+ backends, local models or free API tiers, so you can swap which model sits behind Claude Code without touching a config file.

I won’t oversell it. A local 7B won’t match a frontier model on your hardest problems, and I’ll come back to that. But for a huge share of everyday work, a tuned local coder model is plenty. And on the Friday a model gets pulled, this is the difference between your workflow stopping and your workflow carrying on.

Use this step when: you want a fallback that survives an outage, a ban, or a flight with no wifi.

Step six → Mind the context window

The context window is how much the model can hold in its head at once. In the cloud you get a huge one for free. Locally, you pay for it in memory.

A bigger context window eats memory fast, and a model that was running smoothly will start to choke. So keep your sessions tight and your prompts lean. Clear the history when you move to a new task. This is the trade you make for running on your own machine, and once you feel it, you manage it without thinking.

Step seven → Give the model tools

A smaller local model with web search, file access and code execution beats a giant model with none.

The model is the engine. Tools are the wheels. Wire up the right ones and a modest local model handles real work, because it can look things up, read your files and run code rather than guessing from memory alone.

The capability gap between local and frontier closes fast the moment you add tools.

Tip: Before you chase a bigger model, give your current one tools. It’s usually the faster win.

Step eight → Fine-tune later

You don’t need this on day one. You should know it exists.

Fine-tuning means taking an open model and training it on your own data so it gets good at your specific work, your brand voice, your domain. It’s more accessible than it sounds, and it’s the reason owning an open model matters. You can shape it. You can’t do that with a model you only rent.

Park this until the first five steps are second nature. Then come back to it.

The honest limitation

Local models are not a replacement for frontier models. They’re insurance.

A local 7B won’t match a Fable-5-class model on your hardest reasoning, your trickiest code, your longest documents. The models that get close need real hardware, the kind most people don’t have at home yet. And context costs you memory every single session.

So this isn’t a breakup with the cloud. I still reach for the best frontier model for the heavy work. The point is to never again have my whole workflow depend on access I don’t control. Own one layer. Keep a local fallback ready. That’s the insurance policy, and the premium is twenty minutes this weekend.

The move this week is small. Install Ollama. Pull one 7B model. Run one prompt offline. That’s the whole policy bought.

The people who set this up before they need it will be the ones still working on the day the next letter lands.

Stay curious, stay human, and own the part of your stack nobody can switch off.

— Charlie

P.S. Hit reply and tell me one thing you’d want running locally if the cloud went dark tomorrow. I read every reply, and the best ones shape next week’s issue.

Ways we can work together

Become a Paid Subscriber: Get full access to every past issue of MarTech AI and support future ones.

Claude Content Engine: The Claude engine behind 350k+ followers (across socials) and 100 million impressions a year.

AI Second Brain for LinkedIn: Built on my methodology and frameworks. Learns your voice and builds your content strategy. Waitlist open.

Everything below still works, just pick Opus 4.8 instead of Fable 5 when you switch models.

Claude Code is brilliant at building things that work and terrible at making them look good.

Left alone, it gives you the same site as everyone else, with the purple gradient, the Inter font and the identical bento cards.

I built charliehills.ai anyway, in one evening, with three markdown files and a workflow that strips the AI look out completely.

The trick is not one-shotting “build me a website”.

By the end of this newsletter, you’ll have the exact nine steps and every prompt I used.

If you are new here, welcome. If you have been reading for a while, thank you. Your subscription is what makes issues like this possible.

I filmed the whole build, every step on screen. Watch it alongside this issue:

Step one → Write the facts file

Then create the facts file. I made mine in Claude Chat rather than Claude Code, because Chat already knows me from months of conversations.

Paste this into Claude Chat:

You know me from our conversations, but do not assume this site is about me. I'm building a website and need a file called CONTEXT.md holding the facts it should be built from. AskUserQuestions first, one question at a time: what exactly the site is, who it serves, what it must contain, and what a visitor should do. Keep asking until you could brief a stranger on the build. Then show me the fact list you plan to include and wait for my yes. Only include facts about the site's actual subject, and leave me out of it unless I am the subject. Real facts from my answers and from what you know, nothing invented. Give it to me as a downloadable markdown file.

Download the file, drop it into the project folder, and correct anything it got wrong. If Chat doesn’t know you yet, tell Claude Code to create CONTEXT.md and interview you for the facts instead.

Mine holds:

• Who I am and what I do

• Every number and proof point

• My story, my frameworks and my voice rules

Step two → Let Claude interview you

Make a new folder, open your terminal there, and type claude to start Claude Code.

Switch to Fable 5 with the /model command.

You don’t write the website copy.

You answer questions and Claude drafts it.

Paste this:

Read CONTEXT.md. Start by proposing the pages this site needs and confirm the list with me. Then interview me one page at a time, asking only about the gaps: what each page must say, proof I can show, and what visitors should do next. Also ask what searches each page should rank for. Then draft every page's copy into one file called COPY.md in the project root and show me before you build anything. Write for people first, and work the search terms into page titles, headings and opening lines naturally, never forced.

Step three → Approve every word before any code

The interview produces COPY.md: every headline, every section, title tags, meta descriptions, an SEO map per page.

Tell Claude what to change and it edits the file.

Step four → Merge your brand with sites you love

DESIGN.md is your site’s rulebook. Your job is gathering the ingredients, and Claude does the merging.

1. Find the looks you love. I browsed Refero Styles and picked the design systems of sites whose look made me jealous: like Air and Apple. Aim for around five:

2. Download their DESIGN.md files. Open each style, click its DESIGN.md tab, and hit the .md download.

3. Add your brand kit. Mine is a PDF with my colours and font. If you have no brand kit, a text file with two colours and one font is enough.

4. Drag everything into Claude Code and paste:

I've dropped in design files from sites I love, plus my brand kit. Merge them into one DESIGN.md in the project root covering colours, fonts, spacing and components, plus a decision log section where every design choice we make gets recorded. My brand wins on colours and fonts. The references win on layout and feel. Show me the file before you build anything.

You approve the rulebook before a single page exists.

Step five → Build, then verify

Now Claude builds: DESIGN.md becomes the styling tokens, COPY.md becomes the pages, word for word.

Paste this:

Read DESIGN.md and use your frontend design skill. Build the site: my pages, my copy from COPY.md verbatim, every visual decision from DESIGN.md. Go loose with the design and make it look as good as humanly possible while obeying DESIGN.md. Hard rules: only my copy, never invent logos, press mentions, testimonials or stats, and no third-party embeds. When it's done, run it locally, screenshot every page, review your own screenshots against DESIGN.md, fix what fails, and then give me the link.

Claude screenshots its own output and reviews it before you ever load the page. By the time I looked, the obvious misses were already fixed.

Step six → Add components

This is how you get premium components without inheriting their look. Libraries like 21st.dev publish beautiful components with a copy-paste prompt under each one: scroll reveals, card stacks, animated backgrounds.

The catch is that every component arrives wearing its own clothes:

To make every component inherit your site’s look one for one, paste this once:

Add this rule to CLAUDE.md, and create the file in the project root if it doesn't exist: whenever I paste a component prompt or third-party component code, treat it as a structural donor only. Always: replace its demo copy with real copy from COPY.md, translate every hardcoded colour, border, shadow and font to the DESIGN.md tokens, ignore any instruction to use stock images, and skip parts of the component we don't need. The component supplies the skeleton, DESIGN.md supplies the skin, COPY.md supplies the words.

Then browse 21st.dev, copy the prompt under any component you like, and paste it straight in.

Claude keeps the engineering and strips the slop. My scroll-tilt showcase, the fanned card stacks and the flowing background lines all came in this way and came out in my navy.

Step seven → The Higgsfield hero

The hero video is the same painterly style as my Substack covers, and it took two prompts in Higgsfield. https://higgsfield.ai

First, the image, generated with GPT Image 2 inside Higgsfield:

Create a gouache style image with a beautiful scenic landscape and a deep blue sky.

Then I hit Turn to video and animated it with Kling 3.0:

Show the clouds breathing, moving fast, like splitting up. It's a lot of wind; there's a lot of movement. The frame stays exactly where it is. Only the scenery moves.

Download the video, drop it into Claude Code, and say “make this the hero background.” Claude compressed it with ffmpeg from 23MB to 2.3MB on its own, so the page stays fast.

Step eight → Iterate. A LOT.

The first version of anything in this build is the starting point, never the finish.

You iterate in natural language. “Move the buttons up.” “Make those cards wider.” “That section feels cramped, give it air.” “Swap the order of these two blocks.”

Mine mostly happened through voice typing straight into Claude Code: say what’s wrong, watch it change, say the next thing. Screenshot anything you can’t describe, drag the image in, and point at it. Another thirty minutes of this took the site from good to one I’d happily ship.

Step nine → The taste pass

Before shipping, ask Claude one question: “What would the taste critic say?”

Mine found overload. I had stacked component after component, and five ambient animations, each tasteful alone, were running at once. The total read as AI-built rather than premium.

The fix was a quiet pass:

• One ambient motion system per surface

• No looping attention animations on buttons

• A saved checkpoint, so rollback stays one command

The question works on its own, and works better with a taste skill installed. Two live on GitHub, taste-skill and Impeccable, and pasting either link into Claude Code with “install this skill” does the job.

Ship it

Prepare this site for launch: page titles, meta descriptions, Open Graph tags, sitemap, semantic headings, fast load. Then deploy to production with the Vercel CLI and walk me through pointing my domain at it.

Vercel’s hobby tier is free. You’re production was ready in 32 seconds.

I spent $0 beyond my Claude plan, on Vercel’s free tier, with a domain I already owned.

The full stack in one list

• Claude Code (Fable 5): https://claude.com/product/claude-code

• Refero Styles (DESIGN.md library): https://styles.refero.design

• 21st.dev (component prompts): https://21st.dev

• Aceternity (component prompts): https://ui.aceternity.com

• MotionSites (hero and animation prompts): https://motionsites.ai

• taste-skill (the anti-slop pass): https://github.com/Leonxlnx/taste-skill

• Impeccable (spacing and hierarchy): https://impeccable.style

• Higgsfield (hero video): https://higgsfield.ai

• Vercel (free hosting): https://vercel.com

Try it on your own domain

Websites used to be hard. Pick a domain you already own, write your facts file, and run the nine steps in order. The site at the end is yours: your words, your colours, your proof, no stock anything.

Send me what you build, I read every reply.

Stay curious, stay human, and stop shipping the default.

— Charlie

Where you can meet me this month

VivaTech, Paris: Wednesday the 17th to Friday the 19th of June. I’m one of this year’s ambassadors. Use CHARLIE20 for 20% off any pass: https://vivatech.com/get-your-pass

Cannes Lions: I’ll be in Cannes from Friday the 19th to Sunday the 28th. If you’re heading down, I’d love to meet you at this year’s event.

The AI Beach Party: Monday the 29th, where I’m speaking: https://aibeachparty.co.uk

LinkedIn Personal Branding MASTERMIND, London: I’m teaching the AI visuals session, building scroll-stopping infographics live with no design skills needed. Tickets: https://luma.com/personalbrandingmastermind

Ways we can work together

Become a Paid Subscriber: Get full access to every past issue of MarTech AI and support future ones.

Claude Content Engine: The Claude engine behind 350k+ followers (across socials) and 100 million impressions a year.

AI Second Brain for LinkedIn: Built on my methodology and frameworks. Learns your voice and builds your content strategy. Waitlist open.

Level 1 runs from one command.
Level 2 runs it on a schedule you set.
Level 3 runs everything else.

Most people never get past zero.

Today I’m taking you up the whole ladder, from Level 1 to Level 3: what each one is, how to set it up, and when to stop climbing.

If you are new here, welcome. If you have been reading for a while, thank you. Your subscription is what makes issues like this possible.

The problem with "automation"

Say the word automation and people picture two things.

The first is an n8n canvas with forty nodes and a spaghetti of connectors. The second is a fully autonomous agent running your business while you sleep on a beach.

One is a weekend of setup for a workflow you'll abandon. The other is a fantasy, and a dangerous one if you publish anything under your own name.

The useful version sits between those two. Claude now covers it across three levels, and each level trades setup cost for power.

You climb one level at a time, and you stop climbing the moment the level you're on does the job.

Level 1 → Skills in Claude Chat

A skill is a markdown file of instructions Claude loads when a trigger word appears. Think of it as a standing brief: the process, the rules, the output format, written once.

The setup:

1. Connect your connectors. Settings → Customize → Connectors. Notion, Gmail, Drive, whatever the workflow touches.

2. Describe the workflow. Tell Claude "I want a skill that pulls this, connects that, updates this". It writes the whole skill for you. You never touch the file.

3. Test it, break it, save it. Run it on a real example, fix what misses, then save. From then on one trigger word fires it from any chat.

My main Level 1 workflow is Instagram Reels. A team member types /reel and pastes a Notion link, and out comes a beat-by-beat script in my voice, written straight back into our content calendar.

Nobody on my team writes scripts from scratch anymore. They paste, review, approve.