AI-assisted coding needs smaller loops
Give agent small jobs. Check every result.
ai-tools / workflow / engineering
Big prompt. Big mess.
Small job
Ask for one change. Give files, rules, expected result. Test it. Then next change.
This keeps context small. It also makes bad output easy to find.
Human job
AI can write code. It does not own result. I still decide scope, check tradeoffs, read diff, run tests.
Fast loop beats clever prompt.