AI with no guardrails.

What it is

Your AI coding tool has no memory of your app. Every session starts from zero. It reads a few files, guesses how you do things, and adds a slightly different pattern: a new way to talk to the database, a second way to check who’s signed in, a third place where errors are handled differently. Do that for a few months and the app has seven ways of doing the same thing, and none of them agree.

A project instructions file, a written rulebook, fixes this. It’s a plain-text file at the top level of your project, often called CLAUDE.md, AGENTS.md or .cursorrules, that the tool reads at the start of every session. It gives the tool the rules before it writes a line.

What it looks like from your side

Fixes that work in one screen and not another. The tool “tidying up” something you didn’t ask about. A feature that took an hour last month taking a day now, because the tool has to rediscover how everything fits. A developer who reads the code and says they can’t tell what the conventions are, because there aren’t any.

What it costs if you ignore it

Every session is slower and riskier than the last, and the drift is what turns a messy app into an unfixable one. When the rules for sign-in and data access live in seven places, a security fix has to be made seven times, and the eighth place gets missed. That’s how the problems on the other pages here come back after they’ve been fixed.

How to check it yourself in 30 seconds

How I fix it

I write the rulebook for your app: how it’s organised, the conventions it actually uses, how sign-in and data access must be handled, what must never change without review, and how to run the tests. Then I bring the code in line with it, so the rules describe reality.

From then on every session starts with the rules instead of a guess. It’s part of every Rescue & Refactor, and keeping it current as the app changes is what the retainer is for.