Image generated with AI. Art direction: Grafica.
An AI chat knows only what it is given. Mine start from more, because before they start they read a folder. It is a private repository on GitHub, and I think of it as a brain: the written version of how I work, kept where every AI session can read it. This post is about its shape, and how it learns. What is in it stays private.
The folders
The first entry in its history is dated 7 August 2026. By 19 October it held 1,340 changes, and each folder answers one question an assistant would otherwise ask me.
- Standards: how is this done here? Eighteen files. How a change is classed and rehearsed, how a session closes, what a handoff sheet carries.
- Records: what was decided, and what is still open? Twenty-nine files, one per project or subject. Each decision is dated, with the reason and the options that were ruled out.
- Plans: what comes next? The work, ranked, with who owns each step.
- Lessons: what went wrong before? Twenty-six of them, each tied to a dated mistake.
- Reports: what did each session actually do? Every piece of work ends with a written report that the next step can check.
- Inbox: what has just come back? Sessions drop their reports here and the manager reads them first.
The chats I work in are numbered by brief. This post was drafted under brief 885.
How it learns
It does not learn the way people imagine. Nothing trains a model. It learns because a correction gets written down the same day.
An example. On 9 October I said that featured images should not contain text. That sentence is now a line in the design rules the sessions read, and a script on the repository checks a post card for words in its picture. The next session does not need me to say it again. It reads it, and a check can fail on it.
The same goes for what I reserve to myself. The acts that stay with a person are written down, so a session knows where to stop because the line is on file, not because it remembers being told. Our post How we use AI at Grafica, and the four things it is not allowed to do sets out the four things AI is not allowed to do here.
What «the way I do» means in practice
Not my voice. My judgment, where it can be written: what I check before a page goes to a client, what I refuse to put in a proposal, which numbers I want before I believe a result. When a session meets something the files do not cover, it is supposed to say so, and the answer goes into the files.
Where it falls short
It only knows what has been written. A correction I give in passing and nobody files is lost the next morning.
Files grow. A record that only ever gets longer becomes hard to read, so part of the work is condensing it.
A rule written too broadly misfires. When that happens the rule is narrowed, and the reason goes on file.
If you want one
Start with four files, not a system:
- The instructions you have typed more than once.
- The decisions you have made, each with a date and a reason.
- Your mistakes, each with the rule it taught you.
- A short report template every piece of work ends with.
Then give your AI those files to read before it starts. My earlier post, I stopped writing prompts. I brief a project manager., is about the manager that reads them. If you want to see how this would look for your team, talk to us.
Gráfica built this survey as a WordPress plugin, and builds plugins like it for clients. Tell us what your site needs
