SKILL.md
Risk review 🦆. Break it before users do.
Purpose
Identify highest-impact correctness/reliability/security-adjacent risks and smallest safe mitigations.
Philosophy Guardrails (skill-local)
Inherit shared guardrails from references/GUARDRAILS.md.
Skill-specific delta:
- Risk findings and mitigations only; merge/approval decisions remain with user.
Activation
Use for adversarial review of proposal/change scope.
Method
1. Clarify scope (if missing)
- ask 1-3 targeted clarifying questions when context is incomplete
- state assumptions explicitly when evidence is missing
Required:
- proposal/change scope or runtime path
Optional:
- threat model
- rollback constraints
- compatibility constraints
If threat model/scope is missing, ask one targeted clarifying question.
2. Review for risk
- Inspect trust boundaries and failure classes first.
- Prioritize: security/correctness -> data integrity -> rollback/compat.
- Emit max 3 highest-impact findings.
- Include explicit
Impact+Rollbackin every finding. - Prefer smallest safe mitigation.
- If uncertain assumptions, state assumption in one line and downgrade confidence.
3. Output findings
One line per finding:
<prefix> <path[:line|scope]> — <failure mode>. Impact: <user/data/scope>. Rollback: <blast radius + revert path>. Fix: <smallest safe mitigation>.
Prefixes:
🔴 bug:correctness/security/data-loss🟡 risk:reliability/compat/rollback gap❓ question:missing context blocks judgment
Final line:
totals: <n> findings, <n> questions. coverage: trust-boundary=<checked|partial|missing>; rollback=<checked|partial|missing>.
Boundaries
- No style/naming nits.
- No simplification/duplication ownership (
duck-simplify). - No final PR-thread formatting ownership (
duck-review).