Engineering Craft › AI-Assisted Development
Reviewing AI-Generated Code
Reading and testing AI output as critically as a stranger's PR.
Also known as: AI code review, reviewing AI output, reviewing generated code, verifying AI code
Treat code written by an AI assistant like a pull request from a stranger who is fluent and fast, sometimes wrong, and doesn’t know your project. You’re still responsible for it once you commit it.
Read it all
Don’t accept what you haven’t read. The diff is usually small enough to read line by line; if it’s not, ask for smaller steps.
What to check
- Does it run, and do the tests pass? Then check that the tests really test something. Generated tests may assert whatever the code happens to do.
- Do the things it calls exist? Look for invented functions, methods or options, and for deprecated or version-mismatched APIs (hallucination). Check the official docs.
- Does it solve the actual problem? Confirm it didn’t answer an easier question, or quietly change behavior elsewhere.
- Edge cases and errors: empty input,
None, failures, large inputs. - Security: SQL built from strings, unvalidated input, missing authorization checks, secrets in code, overly broad permissions.
- Fit with your codebase: your conventions, existing helpers (did it reinvent one?), your dependencies. It can’t know unless it was told (agent instructions).
- Unnecessary extras: unrequested rewrites, new dependencies, dead code, over-engineering.
- Licenses and sensitive data if you paste code or data into a tool.
# Plausible, but is `parse_iso` really a method of datetime?
dt = datetime.parse_iso(value) # the real one is datetime.fromisoformat(value)Habits
- Ask for small changes, and review each before the next.
- Run it yourself. “It looks right” isn’t verification.
- Ask the tool to explain parts you don’t follow, then verify that against the docs.
- If you can’t explain it in review, don’t commit it.
Reviewers on your team will review it as yours. “The AI wrote that” isn’t an answer to a review comment. See also self-review.