My AI workflow is boring, and that's the point

People expect that someone who "builds with AI" has some exotic workflow. Mine is almost embarrassingly ordinary, and I think that's why it works.

The loop looks like this: I write a small, precise task. The model drafts. I read every line as if a junior colleague wrote it at 2am — which, functionally, is what happened. Then tests decide whether any of it survives.

Why review still matters

The failure mode I see most in AI-assisted code isn't wrong syntax. It's plausible code that's subtly wrong: an off-by-one that only triggers on the last page of results, an error swallowed so the UI stays happy while data is lost. These pass a glance. They don't pass tests, and they don't pass a slow read.

The rules I actually follow

• never paste code I couldn't explain in an interview
• one task per prompt — kitchen-sink prompts produce kitchen-sink bugs
• the test suite is the referee, not my enthusiasm
• if a fix takes more than two retries, I stop and rethink the problem statement

None of this is clever. That's the point. AI collapsed the cost of producing code, which means the scarce skill moved to judging code. Judgement is boring, repeatable, and valuable.