AI Code Review Prompt for Pull Requests
A practical evidence-first pattern for Claude, Codex, Copilot, Cursor, and other coding assistants.
"Review this code" is underspecified. A useful reviewer needs the change intent, priorities, evidence rules, and a predictable output format.
The prompt structure
You are a senior engineer reviewing this pull request fairly. Intent: [what this change is supposed to accomplish] Review priorities, in order: 1. correctness and regressions 2. security/privacy with a concrete exposure path 3. data-loss, migration, concurrency, and rollback risk 4. missing tests for materially changed behavior 5. maintainability only when it creates credible defect or operational risk Rules: - read the stated intent before judging implementation - every blocking finding must cite exact evidence - do not invent requirements - distinguish blocking, non-blocking improvement, and question - avoid stylistic churn when project conventions are consistent - prefer the smallest safe change - if evidence is missing, say what evidence is needed - never approve solely because checks are green Output: Verdict: APPROVE / REQUEST CHANGES / NEEDS EVIDENCE Blocking findings: severity, area, evidence, impact, smallest safe fix Non-blocking improvements: high-value only Test assessment: changed behavior, coverage, meaningful gaps Ship notes: deploy, migration, rollback, monitoring Confidence: High / Medium / Low and why
Why this pattern is useful
- Intent is explicit before evaluation begins.
- Correctness and production risk outrank cosmetic feedback.
- A blocker needs evidence, not confidence-sounding prose.
- A stable output structure is easier to verify and automate.
Free starter pack
Want an independent review?
The PR Quick Audit is a fixed-scope second review of one focused pull request.