Review Instructions
Severity & Blocking Policy
Block merge: Critical, High findings — [confirm this matches your repo's severity table]
Flag, don't block: Medium findings — [note negotiation criteria, if any]
Optional: Low findings — style/best-practice, does not gate merge
Cite a CWE ID on every Critical/High finding (e.g. CWE-89 SQL Injection, CWE-798 hardcoded credentials)
Always Flag
Unparameterized or string-built SQL queries — require parameterized statements
Unvalidated or unsanitized user input at a trust boundary
Hardcoded secrets, credentials, or tokens, in any format — not just known key patterns
Missing authorization check on a state-changing endpoint
Error responses that leak internal details (stack traces, file paths, query text)
[Add always-flag items specific to your rules that a generic reviewer wouldn't know]
Repo Invariants
[Artifact naming pattern, e.g. docs/adr/NNNN-<topic>.md — flag docs saved elsewhere]
[Enforcement-ladder placement, e.g. "a finding that's already a CI-blocking check is a
duplicate, not a new finding — cite the CI job instead of re-flagging it"]
[Any other invariant a reviewer needs restated every run because it isn't in CLAUDE.md]
Formatting/style nits a formatter already enforces — [name your formatter, e.g. Biome/Ruff/gofmt]
Anything a CI check already blocks — cite the check instead of re-flagging it (see the Enforcement Ladder)
Naming-convention bikeshedding not tied to a correctness or security risk
Project context already stated in CLAUDE.md — restate only what's review-specific
Back to top