agentsclimarketplace
Sign in

Review work

Skill hungv47/meta-skills/skills/meta/review-work

Independent post-implementation review — an agent with no sunk-cost bias checks just-written code or an artifact against its requirements, then resolves the issues found (max 2 rounds). Use to verify a change before shipping, get a second opinion, or check what you missed; auto-triggers for security-sensitive and data-mutation code. Not for code refactoring (use clean-code) or decision analysis (use debate-agents).From its SKILL.md

Install
npx -y skills add hungv47/meta-skills --skill review-work

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 14 stars14 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

6.5 KB, ~1.4k tokens by cl100k_base, as published. Nobody here has run it

Review Chain — Fresh-Eyes Post-Implementation Quality

Independent reviewer + resolver loop for code, artifacts, plans, or shipped work. Reviewer has no implementation context; resolver synthesizes. Routing in routing.yaml; methodology + "no critic on critic" in references/playbook.md.

Core: What would a senior reviewer with no sunk-cost bias catch?

Critical Gates — load first

  1. Reviewer has NO access to implementation reasoning — only the output + requirements.
  2. Resolver sees BOTH original + review — synthesizes, not patches.
  3. Max 2 loops. Not clean after 2 cycles → flag user; may be a design problem review can't fix.
  4. Auto-trigger for critical code — security, auth, crypto, data mutations, money, PII. Don't wait to be asked.
  5. Quality feedback — repeated misses, critic overrides, post-humanmaxxing regressions feed _shared/quality-feedback-protocol.md, not just this report.
  6. Noise-filter before report-write. Every finding passes noise-filter.md: Layer 1 (real-vs-fake per procedures/reviewer.md § Verification rules) then Layer 2 (Accepted / Rejected / Deferred — three report subsections). Accepted clears fix-then-rerun before Verified.

Before Starting

Apply references/_shared/before-starting-check.md. Then:

  • Mode (references/_shared/mode-resolver.md): default standard; auto-escalate to deep for auth/sessions/access-control/payments/financial-data/migrations/bulk-mutations/PII OR diff >500 lines; auto-downgrade to fast for typos/log lines/config tweaks.
  • references/_shared/execution-policy.md — session execution profile (single-vs-multi)
  • Read docs/forsvn/artifacts/meta/specs/*.md + tasks.md if present — enables scope-drift detection per procedures/scope-drift.md.
  • Read docs/forsvn/artifacts/meta/records/learned-rules.md — append to reviewer CONTEXT.

Mode map: fast = generalist, skip resolver if PASS; standard = generalist + resolver loop; deep = 3 specialists parallel OR critic-consensus (non-code).

Artifact Contract

  • Path: docs/forsvn/artifacts/meta/records/[YYYY-MM-DD]-fresh-eyes-<slug>.md (dated, immutable per-run)
  • Lifecycle: snapshot — accumulates, never overwritten
  • Frontmatter: skill, produced_by, version, date, status, mode, rounds, verdict (PASS/FIXED/CRITICAL), provenance. Template: references/report-template.md
  • Sections: Verdict · Issues Found · Input Quality · Scope Drift · Simplifications · Changes Made · Self-Regulation Gate · Reviewer Summary · Resolver Notes · Specialist Verdicts (deep) · Critic Disagreements (critic-consensus)
  • Consumed by: operator audit; future fresh-eyes runs; commit/PR creation (PASS gate)
  • Eval workspace: none — fresh-eyes IS the eval mechanism.

Pre-Dispatch + Execution

  • Pre-dispatch + Warm/Cold Start: references/_shared/pre-dispatch-protocol.md, references/procedures/pre-dispatch.md.
  • Reviewer + resolver built per-use from procedures/reviewer.md, resolver.md — no static agents dir.
  • 8-step loop (identify → spawn → evaluate → resolve → self-regulation gate → loop → report → deliver) + Configuration: procedures/execution.md.

Auto-trigger rules

Run proactively for: security-sensitive code (auth, crypto, access control); data-mutation code (migrations, bulk updates, deletes); complex/uncertain implementations; money or PII.

Do NOT for: trivial changes (typos, config, log lines), "just do it quick" code, read-only ops.

Anti-Patterns + Edge Cases

Read references/anti-patterns.md before: stacking a critic on the reviewer (banned), giving reviewer implementation reasoning, skipping resolver, auto-applying past self-regulation gate, >2 loops, nit-padding. Edge cases (hallucinations, regressions, oversized code, agent failures, architecture reviews) live there too.

Durable Rules (protected)

<!-- SLOW_UPDATE_START --> <!-- No pinned rules yet. Populate via the slow-update workflow (see references/slow-update-fence.md). Each pinned rule must (a) be procedural not instance-specific, (b) be earned from a regression or critic-flagged failure, (c) cite the artifact / decision record that justified pinning. --> <!-- SLOW_UPDATE_END -->

Completion Status

  • DONE — findings resolved or accepted; PASS gate met.
  • DONE_WITH_CONCERNS — non-blocking issues flagged; report names deferred + why.
  • BLOCKED — critical issue (security/data-loss/broken contract) OR self-regulation gate fired; resolver did NOT apply; operator judgment required.
  • NEEDS_CONTEXT — requirements unclear; missing spec/intent/acceptance criteria.

References

What ships with it: 27 files

224.7 KB alongside SKILL.md, 7 of them executable

references/

scripts/

Keep looking

Skills are one crate of 325,949. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.