Move from a list of opinions to findings you can reproduce, test and accept or reject.
Before you start
A small diff, the intended behaviour, relevant interfaces and instructions for running existing tests. Remove secrets.
The workflow
1. Bound the review
Describe what changed and the invariants it must preserve. Include enough surrounding code to understand the inputs, outputs and error paths. Keep unrelated files out of the first pass.
2. Ask for evidence with each finding
Require a concrete failing input or sequence of events, a location in the code, impact and a suggested test. Tell the assistant to label uncertain findings rather than invent context.
3. Reproduce before editing
Run a proposed failing case in a controlled development environment. Reject findings that depend on an imaginary helper or an incorrect assumption about the runtime.
4. Make the smallest verified fix
Apply a focused change, then run the relevant tests and inspect the diff. Human review still owns architecture, authorisation and whether the feature should behave that way.
A starting brief
Replace the bracketed fields. Paste it into ChatGPT, Claude, Gemini or another assistant – or press Fill in.
Review this diff for behavioural defects. Intended behaviour: [DESCRIPTION]. For each finding include the location, concrete failing input, impact and a test that would demonstrate it. If you cannot describe a reproduction, label it speculative. Do not invent the behaviour of code you cannot see. Ignore style preferences unless they cause a defect. DIFF AND CONTEXT: [PASTE]
Make it concrete
Useful finding format: “An empty input reaches this branch; the function returns undefined instead of an empty list. Add an empty-input test.” Confirm it in the actual code.
Check the result
A finding is ready for action when you can reproduce the wrong result and explain the expected result. AI review complements, rather than replaces, tests and ownership.
Keep going
Sources
Product capabilities are based on these official pages. Workflow steps and sample briefs are editorial suggestions, not benchmark results. Checked 25 September 2026.