Every developer who works with AI regularly has a list of corrections they make.

Not a written list. A mental one. The things they fix after every session. The patterns the AI gets wrong consistently. The naming that never quite matches. The component that always ends up doing too much. The state that keeps ending up in the wrong place.

Those corrections are not random. They are consistent. And consistent corrections are rules that have not been written down yet.

You already know what your first rule should be. You have been communicating it to the AI through corrections for months. You just have not written it down where the AI can follow it before it makes the mistake.

Why the first rule is the hardest