Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use before an agent-produced diff is committed, pushed, opened as a PR, merged, landed, or applied to user files when explicit implementation approval is missing. Trigger for review gate, review pack, approve before landing, diff first then land, human approval, merge gate, commit gate, push gate, or PR readiness. Produces a concise review pack, records risks and verification, and blocks landing until a human explicitly approves or approves with required fixes.
.claude/skills/majiayu000-review-gate/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-13 | ✗→✓ | ▲ Improved | -3% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 51% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 14% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 130% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 5% | 0% |
Use this skill as the last-mile safety checkpoint for agent-generated changes. It complements PR review tools; it does not replace code review, CI, or human approval.
Use assets/review-pack-template.md when the user needs a reusable review pack shape.
Run before any of these actions when an agent-generated diff is involved:
that exact implementation path
Read-only review, planning, issue triage, and local exploration do not require this gate unless the next step would land or publish changes.
review pack.
applying a generated patch when the user has not approved that exact action.
surfaces lack review, or approval is ambiguous.
verification commands, or Review Pack risk prompts.
| State | Meaning | Allowed next action | | --- | --- | --- | | draft_pack | Review pack is being assembled. | Inspect diff and verification only. | | needs_fixes | Blocking risks or missing evidence exist. | Patch and rerun the gate. | | awaiting_human | Pack is complete but no human approval exists. | Stop before commit, push, PR, merge, or apply. | | approved | Human explicitly approved the pack in the current thread. | Proceed with the named action only. | | approved_with_fixes | Human approved after specific fixes. | Apply fixes, verify, and record evidence before landing. |
Agents must not self-approve. Prior CI success, a reviewer lane, or a green local test is evidence for the pack, not approval.
pack before landing.
Produce this compact pack:
textreview_gate: - intent: - diff_summary: - files_changed: - driving_skill_or_issue: - risks: - missing_tests_or_verification: - commands_run: - evidence: - open_questions: - approval_needed_for: - decision:
Keep findings ranked by severity. Include exact file paths, PR numbers, issue numbers, command names, and current head SHA when available.
needs_fixes.
innerHTML, eval,shell execution, generated registry, hooks, or high-context files, call that out explicitly.
awaiting_human and ask for the namedaction only.
automatically a merge approval.
the pack before landing.
flowguard should call this gate at landing checkpoints. Queue skills may use a reviewer lane for independent findings, but the Review Gate still records the human-facing pack and approval state.
For GitHub PRs, combine this gate with current remote truth:
For Spellbook changes, the pack usually cites:
bashgit diff --check python3 ./scripts/validate_skills.py --check python3 ./scripts/audit_skill_quality.py skill-name
Use project-specific tests for code changes. If a command cannot run, report the precondition and keep the decision out of approved.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-13 | fail→pass | 12,015 | 5,403 | -55% | 1 | 1 | 0% | 1,935 | 1,886 | -3% | 0 | 0 | — |
case-01 | fail→fail | 4,678 | 14,958 | +220% | 1 | 1 | 0% | 708 | 3,046 | +330% | 0 | 0 | — |
case-02 | fail→pass | 14,404 | 9,552 | -34% | 1 | 1 | 0% | 1,689 | 2,546 | +51% | 0 | 0 | — |
case-03 | fail→pass | 14,553 | 9,100 | -37% | 1 | 1 | 0% | 2,238 | 2,555 | +14% | 0 | 0 | — |
case-04 | fail→pass | 5,492 | 6,251 | +14% | 1 | 1 | 0% | 842 | 1,937 | +130% | 0 | 0 | — |
case-05 | pass→pass | 10,853 | 6,621 | -39% | 1 | 1 | 0% | 1,829 | 2,185 | +19% | 0 | 0 | — |
case-06 | pass→pass | 9,708 | 6,308 | -35% | 1 | 1 | 0% | 1,649 | 1,900 | +15% | 0 | 0 | — |
case-14 | pass→pass | 10,183 | 4,421 | -57% | 1 | 1 | 0% | 1,534 | 1,622 | +6% | 0 | 0 | — |
case-07 | pass→pass | 7,077 | 6,635 | -6% | 1 | 1 | 0% | 972 | 2,015 | +107% | 0 | 0 | — |
case-08 | pass→pass | 11,431 | 6,532 | -43% | 1 | 1 | 0% | 1,640 | 2,040 | +24% | 0 | 0 | — |
case-09 | pass→pass | 15,808 | 8,377 | -47% | 1 | 1 | 0% | 2,626 | 2,371 | -10% | 0 | 0 | — |
case-10 | fail→fail | 4,923 | 6,946 | +41% | 1 | 1 | 0% | 785 | 2,155 | +175% | 0 | 0 | — |
case-11 | fail→pass | 10,496 | 4,467 | -57% | 1 | 1 | 0% | 1,556 | 1,640 | +5% | 0 | 0 | — |
case-12 | pass→pass | 13,662 | 7,105 | -48% | 1 | 1 | 0% | 1,980 | 2,131 | +8% | 0 | 0 | — |
case-15 | pass→pass | 14,721 | 7,665 | -48% | 1 | 1 | 0% | 2,124 | 2,181 | +3% | 0 | 0 | — |
case-16 | pass→pass | 9,610 | 7,444 | -23% | 1 | 1 | 0% | 1,746 | 2,327 | +33% | 0 | 0 | — |
case-17 | pass→pass | 2,573 | 2,338 | -9% | 1 | 1 | 0% | 391 | 1,283 | +228% | 0 | 0 | — |
case-18 | fail→fail | 3,928 | 5,932 | +51% | 1 | 1 | 0% | 523 | 1,779 | +240% | 0 | 0 | — |
case-19 | pass→fail | 5,015 | 8,697 | +73% | 1 | 1 | 0% | 939 | 2,452 | +161% | 0 | 0 | — |
case-20 | fail→pass | 9,367 | 6,630 | -29% | 1 | 1 | 0% | 1,486 | 2,080 | +40% | 0 | 0 | — |
case-21 | pass→pass | 12,520 | 10,145 | -19% | 1 | 1 | 0% | 1,946 | 2,579 | +33% | 0 | 0 | — |
case-22 | fail→pass | 16,183 | 2,669 | -84% | 1 | 1 | 0% | 2,373 | 1,342 | -43% | 0 | 0 | — |
case-23 | fail→pass | 4,178 | 2,963 | -29% | 1 | 1 | 0% | 577 | 1,379 | +139% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted. The headline lift of +30 percentage points is the difference between those two pass rates over the 23 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.