On my Arabic-language channel I review AI coding tools with one constraint: the tool does a real task in a real repo, and I show the full diff. No jump cuts over the part where it hallucinates an API.
The scorecard
- Correctness: does the change compile and pass tests on the first try?
- Scope discipline: did it touch only the files the task needed?
- Explainability: can it say why, not just what?
- Reversibility: is the diff small enough to review in five minutes?
The pattern across every review: AI is a fast junior: excellent at boilerplate and migrations, unreliable at architecture. Keep it inside guardrails (tests, CI, small diffs) and it earns its keep.