Understanding AI

The folder that writes like you

The fix for writing that doesn't sound like you isn't a better prompt. It's a folder.

Build Yours Now

This is the whole job in one paste. Put it into Claude or ChatGPT exactly as written. It asks you for a couple of things you have actually written and works out your voice from those rather than asking you to describe it, because nobody can describe their own writing accurately. Then it interviews you about your business and the formats you write most, and hands back every file finished and named, ready to save into a folder. About twenty minutes, most of it answering questions.

Paste this into your AI

I want you to build my writing folder with me. It's a small set of files I'll put in front of you before you write anything, so that what comes back sounds like me instead of sounding like AI. I want to walk away with every file finished and ready to save. Work through this in four stages. Don't start a stage until we've finished the one before it, and ask me one question at a time throughout. Stage one, my voice. Ask me to paste in two or three things I've actually written. Emails, posts, a report, a message to a client, anything that went out in my own words. Then read them properly and work out how I write from the samples themselves. Sentence length and rhythm, how formal I am, how I open and how I sign off, the words and phrases I reach for, the words I never use, punctuation habits, British or American spelling, whether there's humour in it and what kind. Don't ask me to describe my own voice, because nobody can do that accurately about themselves. Tell me what you found, quote the lines that showed you, and let me correct you. Stage two, the context. Interview me about who I am, what my business does, who I'm writing for and what I want them to do about it. Ask about the things I probably haven't thought of. If I give you a vague answer, push back and ask for the specific version. Only ask about things that would genuinely change how a piece gets written. Stage three, the formats. Ask me which two or three things I write most often. For each one, ask me how it's usually structured, what a good one does, and what a bad one does. Keep going until you could write one without me in the room. Stage four, write the files. Output all of them in full, clearly named, ready for me to copy out and save. 1. CLAUDE.md, the brief. This is who you are when you write for me and how you behave. My voice turned into instructions, built from what you found in my samples rather than from adjectives, with a short always list and a short never list. Include a line telling you to read the context file first and to follow the relevant skill file when the piece is a known format. 2. context.md, the background. Plain factual statements about me, my business and my reader. Not instructions to you, just what's true. 3. One file for each format I named, each saved under a skills folder with the format as its filename, so a newsletter skill would be skills/newsletter.md. The steps in the order they happen, and what good looks like at each step. Write them as finished files with nothing left for me to fill in. No square brackets, no "add your own here", no notes to myself. Use my actual words from the samples wherever they fit. Keep each file short enough to read in a minute, because a long file hides the lines that matter. Last thing, and be honest about it. Tell me the one or two lines across all the files you were least sure about and why you weren't sure. That's where I'll start when the first draft comes back not quite right. One question at a time, and start with stage one.

One thing worth doing properly. When it asks for samples, give it real ones that went out, not the polished version you wish you wrote. The whole point is that it copies how you actually write, and a tidied up sample gets you a tidied up voice that still isn't yours.

Where The Folder Lives

Copy each file out of the chat, save them with the names it gave them, and put them in one folder on your computer. That folder is the master. Everything below is just a way of getting it in front of the AI before it starts writing.

In the Claude desktop app. This is the one in the video and the one I use. Save the folder somewhere sensible, then in the chat box type @ and pick the folder. Claude reads what is in it. Then dictate or type your braindump, the rambling version, everything you want to say in whatever order it comes out. What comes back is written in your voice and it is ready to use. Dictating beats typing here because you want as much raw material in there as possible, and you talk faster than you type.

In a Claude Project. Make a Project, add context.md and your skill files to the Project knowledge, and paste the text of CLAUDE.md into the Project instructions box. Every chat you start inside that Project has all of it already. Same result, no folder on your computer.

In ChatGPT. Same shape. A Project, the files added to it, and the CLAUDE.md text in the instructions. If you would rather not use a Project, upload the files at the start of a chat and it works for that one conversation.

Anywhere else. Whatever tool your work has given you, the pattern holds even when the menu names do not. If there is somewhere to attach files or add knowledge, that is where they go. If there is not, paste the text in at the top of the chat before your first message. That is not a worse version, it is the same information arriving by hand.

One honest note. Nothing here syncs. The AI is reading the copy you gave it, not watching your business, so when your voice shifts or the business moves you go and edit the files yourself. It takes a minute and it is the only maintenance there is.

What's Actually In It

Nothing clever. It is exactly what you would hand a new starter if you asked them to do this job for you. Three things, and you would not think twice about giving a person any of them.

The brief, who you hired

CLAUDE.md

One document that gets read alongside every message you send. Not the task, the task changes every time. This is who this writer is and how they behave, so it shows up the same way every day instead of improvising a new personality each morning. Your voice, written down as instructions, plus the things they must always do and the things they must never do.

The context, what they need to know about

context.md

Who you are, what your business actually does, who you are writing for. A new starter can have perfect manners and still write something wrong, because nobody has told them anything about the place. This is the file that makes the writing accurate rather than just well mannered, and it is plain statements of fact, not instructions.

The skills, the step by step per format

skills/

One short file for each type of thing you write. A LinkedIn article, a LinkedIn post, a newsletter. The steps in the order they happen and what a good one does at each step, because a newsletter and a long article are not the same job even in the same voice. You only need files for the formats you actually write, so most people start with two.

Mine, The One In The Video

This is the actual folder I opened on screen. Five files, none of them long, and every LinkedIn post, article and newsletter I publish goes through it.

The folder

writing style/ CLAUDE.md the brief, who you hired context.md what they need to know about skills/ linkedin-article.md step by step for a long piece linkedin-post.md step by step for a feed post newsletter.md step by step for an email

The brief is the shortest file and it does the most work. Here is the end of mine, the two lists, which is the part that stops the drafts drifting back towards sounding like everyone else.

From my CLAUDE.md

## Always - Coin a name for the central idea and build the piece around it - "the penny drops" for a realisation - Sign off "Thanks for reading," then "Oli" on its own line ## Never - Emojis. Anywhere. No exceptions. - Corporate vocabulary. No leverage, stakeholders, or transformation journey. - "game changer", "unlock", "supercharge" - A hard sell. The most I ever do is invite a message.

Notice how specific the nevers are. Not "avoid corporate language", but the actual words, because a vague rule gets ignored and a named word gets caught. Most of those lines got added one at a time, each after a draft came back with the thing in it.

One Honest Bit

The first version of these files will not be perfect and you will still spend time editing what comes back. That is not the thing going wrong, that is the first week.

The useful part is what the editing tells you. When a draft comes back and something in it is not quite you, that line is the folder showing you what it was never told. Go and put it in the file, one sentence, and move on. Do that for a couple of weeks and the edits get smaller and smaller, which is exactly how it goes with a new starter. Your job is to keep correcting until they are good enough to do the work on their own, and then it is genuinely off your plate.

And the honest version of the promise. This makes the writing sound like you. It does not make it right. The facts, the claims and the numbers are still yours to check before anything goes out with your name on it, same as work from anyone who started last week.

Where This Goes Next

What you have just built is an agent, in the ordinary sense of the word. A brief, some context, and a job it knows how to do. An AI agent is a chat with a job is the one pager on what that actually means, and it takes about three minutes.

The two files in your folder each have a page of their own if you want to go deeper on one. The brief is your agent's personality covers the five things a good brief contains. The context is what your agent actually knows covers the background file and why stuffing everything into it makes the answers worse.

And if you build things as well as write them, the CLAUDE.md guide is the deeper version of that first file, written for Claude Code rather than for writing.