Ask Gemini to review a diff for correctness, security, and clarity, rank findings by severity with a concrete failure scenario each, and skip style nits unless they hide a bug.
The prompt
Review this change for correctness, security, and clarity. For each finding: give a severity (blocker / major / minor), the specific failure scenario, and the smallest fix. Skip pure style nits unless they hide a bug. If the change looks safe, say so plainly. Diff: [DIFF].
Replace the variables
[DIFF]The diff or the changed files.
Which model to use
- Gemini 3.1 Pro — Severity judgement and failure scenarios need careful reasoning.
Example input
- [DIFF] = a pull request touching auth middleware.
Example output
We only publish example outputs from real model runs. This template has not been formally tested yet, so no output is shown. Run it yourself with the input above.
Why this structure works
- Severity + failure scenario filters out noise and false alarms.
- Permission to say "looks safe" reduces invented problems.
Common failures & fixes
Every line gets a comment.
Fix: Cap it: "Report at most the 5 most important findings."
Known limitations
- A model review complements, not replaces, human review and tests.