Image generated with AI. Art direction: Grafica.
I no longer write a prompt for each task. I talk to one AI project manager. It coordinates the work through a written memory of how Grafica operates, sends each task to the chat for the right project, and keeps me from repeating instructions. My part is to steer it in the right direction from experience.
A prompt lasts as long as its chat
A prompt is a request typed into a chat. The next chat starts without it, so the same instruction gets typed again, a little differently each time.
A brief is different. Here, every task is one numbered block of text. It names the project it belongs to, what is to be done, what may not be done, and how the result will be reported and checked. The project manager writes it from what I ask and sends it to a chat that works on that one project and nothing else. The brief that drafted this post was number 758, and one of its lines said that nothing may be published before I read it.
What the manager reads first
The manager does not start from a blank chat. Before it writes a brief, it reads a private repository of rules, records and lessons. When I correct something, the correction is written into that repository, so the next chat starts from it instead of from me. On 12 October 2026 the repository held 17 files of standards, 27 records and 25 lessons, and the first entry in its history is dated 7 August 2026.
What is in it stays private.
Where it stops
Some acts stay with a person, whatever the tool can do. They are listed in How we use AI at Grafica, and the four things it is not allowed to do. The manager knows where that line is because the line is written down, not because it remembers being told. The speed this buys is real, and so is the need for someone to steer, as our case study of a site whose first build took one day shows.
How a result is checked
A chat saying «done» is not evidence. Each brief ends with a written report, and the manager checks the report’s claims against the thing itself, the live page, the file or the setting, before anything counts as done. When the report and the page disagree, the page wins.
If you want to work this way
- Write down the instructions you repeat. Anything you have typed twice belongs in a file, not in a chat.
- Keep one chat per project. The chat that drafts your newsletter should not be fixing your website.
- Write the brief before the work starts: what must be true at the end, and how you will check it.
- Write down the decisions that stay yours, so the AI knows where to stop.
- Check the result against the real thing, not against the report about it.
This is how we run our own work. If you want to see what it would look like on a project of yours, talk to us.
Gráfica built this survey as a WordPress plugin, and builds plugins like it for clients. Tell us what your site needs
