Evaluating AI output like a code reviewer, not a fan

The biggest productivity trap in AI-assisted development is approval. The code looks right, so your brain says yes and moves on. I've lost hours to bugs that survived exactly this moment, so I built myself a checklist and I follow it even when the code looks perfect.

The checklist

Correctness against the actual requirement. Not "does it run" — does it do the thing, including the annoying details I specified? AI is a yes-machine; it will implement a wrong requirement just as beautifully as a right one.

The edges. Empty inputs. One item. Ten thousand items. Unicode in that name field. Timezones. Null that shouldn't exist but does. Machine-written code tends to handle the happy path beautifully and the edges generically.

The quiet sins. Swallowed exceptions. Cache that never invalidates. A query inside a loop. A dependency added for one function. None of these fail the build. All of them fail production.

Ownership. If I can't explain the code line by line, I don't ship it. This isn't pride — it's maintenance. Six months from now, the person debugging at midnight is me, and "the AI wrote it" fixes nothing.

The meta-skill

Review speed is a skill now, the way reading stack traces was a skill. Slow down exactly where the machine is confident. Confidence in output is not correlated with correctness, and learning to feel that gap is the actual job.