Understanding AI

You became your AI's assistant

The fix is boring and it changes everything. One folder structure you actually work out of.

Build Yours Now

This is the whole job in one paste. Put it into Claude or ChatGPT exactly as written. It asks you about the parts of your life and work where an assistant would actually help, picks two or three to start with, then interviews you for each one the way you would brief a new starter. What comes back is the full structure and every starting file written out, ready to save. Half an hour, most of it answering questions.

Paste this into your AI

I want you to build my folder structure with me. It's the set of folders and files I'll keep on my computer and open in front of an AI, so that when I ask for something you already know who I am and how my work runs, instead of me explaining it all again every time. I want to walk away with the whole structure and every starting file written in full, ready to save. Work through this in four stages. Don't move on to the next stage until we've finished the one before it, and ask me one question at a time the whole way through. Short questions, plain English, no long forms to fill in. Stage one, work out my branches. Ask me about the parts of my life and work where an assistant who already knew everything would genuinely save me time. Don't assume what they are, find out. Some people have a business, some have a job, some have clients, side projects, a house and a family to run, something they write every week. Keep asking until you have the real picture, then suggest two or three branches to start with and let me change them. Two or three, not ten, because I have to actually fill these in. Stage two, the top of the structure. Interview me for the things any AI should know before it helps me with anything at all. Who I am, what I do, what I run, where I am, who is around me, and how I like things done. The standing facts that don't change from one task to the next. If I give you a vague answer, ask me for the specific version. Stage three, one branch at a time. For each branch, interview me the way you would brief a new starter hired to help me with that part of my life. How it works, who is involved and what they are called, how things get done and in what order, what is always true, and what someone would get wrong if they were guessing. Finish one branch properly before you start the next one. Stage four, write it. Output the full structure as a file tree, then write out every file in that tree in full, in the order they appear. A context file at the top for the standing information, and one context file inside each branch. Nothing left for me to fill in. No square brackets, no "add your own here", no notes to myself. Keep every file short enough to read in a minute, because a long file buries the lines that matter, and use my own words from the interview wherever they fit. Then two last things, in this order. Tell me the rule that keeps this alive. Every time I catch myself explaining something to an AI that I have explained before, that sentence goes into the file it belongs in. That is the whole maintenance job, and it is how the structure ends up knowing me. Then be honest with me. Tell me the one or two areas you were least sure about, where you were filling gaps because my answers were thin, and what you would ask me about next time. That is where I start when something comes back wrong. One question at a time, and start with stage one.

One thing worth doing properly. When it asks what a new starter would need to know for a branch, answer it like you are actually handing the work over, including the bits you think are obvious. The obvious bits are the ones you have been retyping into chats for months without noticing.

Setting It Up

Copy the files out of the chat and save them into folders that match the tree it gave you, on your computer or in your Drive, wherever your work already lives. Then the only rule that matters is that you open the folder at the top and work from there, rather than starting a fresh chat from nowhere every time.

In the Claude desktop app. Type @ in the chat box and pick the top folder. It can see everything underneath it, so you ask your question there and it goes and finds the branch it needs.

In Claude Code. Open the folder and start. There is nothing to attach, it reads what is around it, which is why I do most of this there.

In Claude Projects or ChatGPT Projects. There are no nested folders, so make one Project per branch and put the top level context file in each of them alongside that branch's files. You lose the automatic routing and pick the Project yourself, which is a small tax for not having to explain yourself twice.

One honest note. Nothing here syncs. The files only know what you put in them, so they go stale on their own unless you keep them current, and the way you keep them current is the re-explaining rule further down.

Why This Works

At some point without noticing, you became the assistant. You fetch its answers out of one app and carry them into another, you paste the same background in at the start of every chat, and you do the remembering for a thing that is supposed to remember for you. That is not the AI being weak, it is the AI arriving to every conversation with nothing.

A structure flips it. Everything it would have asked you for is already sitting there, filed where it belongs, so the fetching and the carrying stop being your job.

The shape of it

my work/ context.md who I am, what I run, how I like things done client work/ context.md the clients, how the work runs, what good looks like personal work/ context.md the house, the admin, the money, the diary content/ context.md what I make, who it is for, where it goes writing style/ the writing folder, one branch of its own

Yours will have different branches and that is the point. The shape holds whatever your life is actually made of.

The folder at the top

context.md

The one you open and work out of. It holds who you are, what businesses or jobs you run, and the general context any AI needs before it can be useful about anything. This is the stuff you currently retype at the start of every chat, written down once. It is short, it is factual, and it changes about twice a year.

The branches, one per part of your life

client work/, personal work/

Underneath it, a folder for each area an assistant could actually help with. Client work, personal admin, the thing you write, whatever is true for you. Each one holds the documentation that lets AI be genuinely helpful there, which is the same pack you would hand a new starter hired for that part of your life. How it works, who is involved, how things get done. Start with two or three.

What happens when you ask

the routing

You ask your question in the folder at the top, and you are not starting from nothing. It reads the standing context, works out which branch the question belongs to, and goes and gets what it needs from there. You stop being the one who fetches the right document and pastes it in, because the structure already tells it where everything is.

Then Connect Your Apps

The structure on its own gets you an AI that knows things. Connecting your apps to it gets you one that does things. Same question, but now the answer lands as a draft sitting in your Gmail, or an event on your calendar, instead of a block of text you have to go and put somewhere yourself.

That is the last bit of assistant work you are still doing, and it is worth doing in that order. Context first, so that when it can act, it acts on something that knows your world. Tools are how your AI stops being useless has the two minute setup for Claude and ChatGPT, which apps to connect first, and the four I would never connect.

One Honest Bit

You do not build this in a day. You set up the initial structure, you get the obvious context in, and there will be plenty you miss, because nobody can sit down and list everything they know about their own work.

So the gaps show themselves in use. Every time you catch yourself re-explaining something in a chat, that is the structure telling you what it was never told. Put the sentence in the file it belongs in and carry on. That is the whole maintenance job, it takes a minute, and after a few weeks the re-explaining mostly stops. In a way you are getting your system to learn from you.

And the honest version of the promise. A structure makes AI helpful, it does not make it right. Everything that comes back still gets checked before you send it or act on it, same as work from anyone who started last week.

Where This Goes Next

If you want a first branch to build properly, build the writing one. The folder that writes like you is the same idea at branch level, with a prompt that reads a couple of things you have actually written and works out your voice from those. It drops straight into the tree above.

For what goes inside a context file and why stuffing the entire filing cabinet in makes the answers worse, the context is what your agent actually knows. For the connections, tools are how your AI stops being useless.

And if the word agent has been floating past you for a year without anyone defining it, an AI agent is a chat with a job is the three minute version. What you have just built is the ecosystem those agents work out of.