Do you trust your AI to do what you say?

I've been experimenting with a workflow where I give an AI a predetermined test suite and ask it to write only enough code to pass those tests. Think of it as "building the app from the tests."

What's surprised me is how hard it is to stop the model from doing extra work. Smaller models especially seem to ignore constraints, they'll refactor, add features, or change things that weren't requested.

For those of you using Claude Code (or similar tools), how much do you rely on .claude files and hooks to enforce behavior versus just prompting the model?

If you don't use hooks, what's your approach for keeping the AI inside the boundaries you want? Do you have a workflow that's been reliable?