The useful result
Review one AI-written change with seven questions and decide to accept, change or reject it.
- Review one AI-written change with seven questions and decide to accept, change or reject it.
An assistant returns two hundred changed lines and says “done”. Accepting without review is how unexpected files, new packages and deleted tests slip in. Reading every line is not realistic for everyone either.
A fixed set of questions makes the review short and repeatable. The questions below look at what changed, what was not requested, what was removed, how data and secrets are handled, and whether you can explain the change. OWASP’s guidance for language-model applications is a useful reminder to treat model output as untrusted until checked.
The examples below are illustrative exercises, not recorded results from any tool. Sources were checked on 24 October 2026; menus, limits and prices change, so confirm details on the linked pages.
Before you start
- A Git repository with an AI-written change not yet committed
- A terminal or Git app that shows a diff
- The original request you gave the assistant
Review evidence
Documentation-based review · 2026-10-24
Primary documentation and official guidance checked on 24 October 2026. Examples, checks and exercises are original illustrations, not recorded tool tests. Cover artwork is a conceptual illustration.
Read the size of the change before the content
Run a summary of the diff to see which files changed and how many lines moved. Compare the list with your request. Files you did not expect to change, a change much larger than the task or deleted tests are your first red flags.
Check your result
You can say how many files changed and whether each one was needed for your request.
Ask the assistant to explain each change, then verify the claims
Paste the diff and ask for a plain-language explanation of each hunk and a list of risks. Treat the explanation as a claim to check against the code. Note any new dependency, any removed code and any place that touches secrets, personal data or file paths.
Explain this diff to a teammate. For each changed file give: what changed, why it was needed for the request “[your request]”, and any risk. Then answer: 1) Which files were not needed? 2) Any new dependencies? 3) Anything deleted? 4) Any handling of secrets or personal data? 5) What happens with empty, huge or invalid input? Do not say “looks fine” without evidence from the code.
Diff:
[paste diff]Example / expected result
Illustrative example; not a recorded model result.
“package.json gained a new date library that the request did not need.”
Check your result
Every hunk has an explanation you have checked against the code, and new dependencies are listed.
Run the checks, try three odd inputs and apply the seventh question
Run your tests, then try an empty value, a very long value and an invalid value by hand. The seventh question decides the outcome: can you explain the change to someone else in two sentences? If not, ask for a simpler version or reject it.
Check your result
Tests pass, the three odd inputs behave sensibly, and you can explain the change in two sentences.
When it goes wrong
An assistant explaining its own code can sound confident and still be wrong, so verify against the diff. Security-sensitive changes, such as login or payments, deserve a human expert. Review smaller changes more often rather than one huge one.
Make it your own
Review your next AI-generated change with the seven questions and note which one found something. Keep a running list of the questions that catch the most problems.
Useful next resources
Sources and review
Last checked Oct 24, 2026. Features and account limits can change. Verify important results against your original material.
Did this help you complete your task?
Tell us what needs improving. No account or private output required.
Report incorrect information or a broken linkGenius Forges Editorial
Practical AI lessons with checked sources, labelled examples and human verification.
