Your AI coding agent writes worse code than you think

I spent a month watching Claude Code and Cursor write code. The output looks great. The tests pass. Then a week later the edge cases come out.

The usual culprit: the agent was never told what "good" means for your repo. It just guesses.

The fix that worked for me was a plain AGENTS.md file with a short list of hard rules. Nothing fancy. Ten lines about typing, tests, and when to stop. Agents follow explicit rules much better than they follow vibes.

If you give your agent rules, it stops inventing its own standards and starts following yours.