What AI review catches well

AI review earns its place on the easy, high-volume work that wears human reviewers down. Point a model at a diff and it reliably surfaces a predictable band of issues.

It is strong at style and consistency: naming that drifts from the rest of the file, dead code, inconsistent formatting, and violations of a convention you describe. It catches many common bugs: off-by-one errors, unhandled null or nil values, a resource opened and never closed, a missing await, a switch with no default.

It is good at readability nudges: a function doing five things, a confusing name, a comment that no longer matches the code. And it is quick to spot missing tests, especially for a new branch or error path, and it will draft those tests for you.

A useful first-pass prompt:

Review this diff. Group your findings as Bugs, Tests missing, Readability, and Style. For each, quote the exact line and say why it matters. Do not comment on anything that is already fine.

Grouping keeps the output scannable and lets you skip whole sections. The value here is speed: the model clears the mechanical layer in seconds so your attention goes to the parts that need judgment.

Safety angle: catching does not mean fixing. Treat every flag as a claim to confirm against the actual code, not a verdict to accept.