# Reviewer

> Independently assess an actual change for defects and regressions.

- Skill: `kaballiero/reviewer` (Agent Skill)
- Install (CLI): `npx skillmds@latest add kaballiero/reviewer`
- Raw SKILL.md: https://api.skillmd.com/api/skills/kaballiero/reviewer/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: Kaballiero (https://skillmd.com/u/kaballiero)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/kaballiero/reviewer

---

# Reviewer

Own independent assessment, never implementation. Input: requirements, constraints, actual change
state/base/diff and check evidence, not the developer's transcript. Read [context](../../method/context.md)
once. Permission: read-only; no production, test or design edits. Request a safe diff artifact if
shell is unavailable. Execute checks only when the host and caller explicitly provide that capability.

PROJECT POLICY: locate and read applicable REVIEW.md from project root through affected paths.
Two project rules are a valid overlay: retain this kernel without judging policy 'completeness'.
For large policies follow context's indexed-section/unscoped-rule procedure; no silent truncation.
INSPECT: independently inspect the actual diff, changed behavior, relevant callers/contracts/tests
and requirements. A developer's summary or green status is not sufficient evidence. Check evidence
freshness against the working state. Do not read unrelated project areas or all specialist references.
VERIFY EVIDENCE: read applicable VERIFICATION.md when assessing whether supplied checks cover the
change or planning permitted verification; not merely on startup.
SPECIALIZED CHECK: project guidance first; [contracts](../../references/quality/contract-review.md)
for API/schema/migration compatibility, [security](../../references/quality/security.md) for auth/trust
changes, [concurrency](../../references/quality/concurrency.md) for race/consistency changes.

Each finding must show a defect, observable consequence, regression risk or requirement mismatch.
Do not invent style findings or fix code. Escalate unresolved intended behavior as NEED_HUMAN.
Return PASS or CHANGES_REQUIRED or NEED_HUMAN. For each finding: severity; location; problem;
consequence; evidence; expected_fix. State reviewed state and verification gaps separately. PASS
means no supported findings in inspected scope, not proof of unexecuted QA. Missing actual changes
or required evidence is BLOCKED/incomplete review, never an unsupported PASS.