| name | code-review |
| description | Guidelines for AI code review — finding real bugs, not style nits. TRIGGER when reviewing code, diffs, PRs, or implementations. |
Code Review
Find bugs that would break production. Not improvements. Not style.
A Finding Must Be
- Introduced in this change — not pre-existing
- Provably impactful — name the scenario/input that triggers it. Speculation is not a finding.
- Actionable — a discrete fix exists
Severity
- P0 — Blocks merge. Breaks production, loses data, security hole.
- P1 — Should fix before merge. Realistic conditions trigger it.
- P2 — Worth fixing eventually. Unlikely in practice.
- P3 — Nit. Do not report.
Don't inflate. P2 called P0 erodes trust.
Verdict
Every review ends with exactly one:
- LGTM — no P0/P1. Ship it.
- FAIL — P0/P1 found. Must fix.
LGTM = short. Don't pad with praise.
Process
- Read
plan.md if it exists
- Read changed files (use
read tool, don't guess from context)
- Run tests — failure is P0
- If shared state changed, run race detector
Output
## Findings
### [P0] Title
**File**: path.go:42-48
**Issue**: what's wrong and what triggers it
**Fix**: what to change
## Verdict
LGTM / FAIL
