How I Steer My AI With Config Files

Anyone who develops with AI runs into the same thing fast: an AI has no memory between sessions. Open a new conversation and it starts from zero. No knowledge of your project, your preferences, your agreements, your way of working.

That is why I do the same thing on every project I build with AI. I capture the context in files the AI reads every session. Playitsmart was no different. It is not a trick I found along the way, it is how I work.

It is about context

The core is this: an AI is only as good as the context it has. Give it nothing, and it delivers generic work. Give it the right context, and it works the way your project needs.

That context I capture in config files. Not loose notes, but files in the repo, under version control, growing along with the project.

Sometimes that documentation is almost as large as the code itself. That sounds like excess, but it is not. This documentation is not written for a human reader. It is aimed at the AI, so that at any moment it has the context to deliver good work. You make what works.

Two files

For the thinking work I use a file with project instructions. It describes the role of the AI, the way of working together, the hard rules, and the context of the project. That file is read at the start of every session. The AI immediately knows who it is, what we are building and what the agreements are.

For the code there is a second file, a set of rules Cursor follows while writing. Code style, patterns, conventions. Cursor reads it along with every instruction.

What goes in

Not just style. The most important rules are about safety. One of them: code changes that touch money get extra caution, correctness over speed. Another: a certain setting may never be permanently on for the production servers, because that once caused two weeks of silent errors. That lesson is now in the file. The next session can no longer make it.

That is the difference between a rule in your head and a rule in a file. The rule in your head holds as long as you remember it. The rule in the file holds always, even in a session where you are not thinking about it.

It is like a handbook for a new colleague

Compare it to onboarding someone. You can tell every new colleague everything verbally and hope it sticks. Or you write a handbook that everyone reads before they start.

An AI is, in a sense, a new colleague every session. Competent, but without knowledge of your project. A good instructions file is the handbook. The difference between an AI that delivers generic work and an AI that works the way your project asks sits largely in that one file.

The takeaway

Capturing context is not a side task of developing with AI. It is the main task. An AI without context guesses. An AI with good context works.

So write it down. Not for yourself, but for the AI. Sometimes that is almost as much text as code. That is not waste, that is the work.