Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Convert accepted repository rules into the smallest repository-native mechanical validation with positive and negative proof and precise enforcement-level reporting. Use for requests to enforce architecture, reliability, security, or quality boundaries; prevent a documented violation from recurring; add structural guards; or turn accepted rules into validation. Do not use to infer or invent policy from conventions, code patterns, tests, defaults, or undocumented preferences.
.claude/skills/hoangnb24-encode-invariant/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | -38% | 0% |
| case-20 | ✓→✗ | ▼ Worse | -33% | 0% |
| case-21 | ✓→✗ | ▼ Worse | -53% | 0% |
| case-22 | ✓→✗ | ▼ Worse | 74% | 0% |
| case-17 | ✓→✓ | ＝ Same ✓ | 0% | 0% |
Turn an accepted rule into a focused guard without creating new product policy.
Read applicable AGENTS.md, docs/WORKFLOW.md, and the complete encoding pattern. Locate the repository's product or technical authority and its native validation owner.
Cite the accepted source and restate the exact scope, allowed behavior, forbidden behavior, and authorized exceptions. Stop before edits if authority is absent or materially ambiguous.
Do not promote conventions, code patterns, tests, defaults, tool behavior, or undocumented preferences into policy. An existing check without accepted authority is a mismatch to report, not authority to expand.
Reuse the repository's existing test, build, task, lint, scan, or validation owner. Choose the lowest deterministic layer that sees the complete accepted scope. Avoid a parallel framework or duplicated source of truth.
Make failures actionable: name the violating item, the broken rule, the authority source, and a concrete compliant next action.
Implement only the authorized boundary. Add or run:
diagnostic.
Use recoverable fixtures or test mutations for negative proof. Preserve unrelated work and never leave a deliberate violation in product files.
Inspect the checked-in validation and CI paths. Do not install hooks, choose a CI provider, change merge policy, or mutate external branch protection unless the user separately authorizes that action.
Report each level independently:
verified; and
Source presence is not execution; CI presence is not a required check; a green check is not proof of merge blocking.
Return accepted authority, encoded scope, changed owner, actionable diagnostic, positive and negative results, the four enforcement levels, and unresolved risks or authority gaps.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 6,927 | 14,596 | +111% | 1 | 1 | 0% | 392 | 836 | +113% | 0 | 0 | — |
case-02 | fail→fail | 7,001 | 6,790 | -3% | 1 | 1 | 0% | 454 | 762 | +68% | 0 | 0 | — |
case-03 | fail→fail | 15,299 | 21,205 | +39% | 1 | 1 | 0% | 574 | 772 | +34% | 0 | 0 | — |
case-04 | fail→pass | 41,854 | 5,697 | -86% | 1 | 1 | 0% | 2,131 | 1,328 | -38% | 0 | 0 | — |
case-05 | fail→fail | 21,782 | 9,118 | -58% | 1 | 1 | 0% | 3,322 | 863 | -74% | 0 | 0 | — |
case-06 | fail→fail | 2,426 | 9,081 | +274% | 1 | 1 | 0% | 201 | 1,041 | +418% | 0 | 0 | — |
case-07 | fail→fail | 12,851 | 9,513 | -26% | 1 | 1 | 0% | 2,366 | 975 | -59% | 0 | 0 | — |
case-08 | fail→fail | 18,397 | 6,697 | -64% | 1 | 1 | 0% | 214 | 841 | +293% | 0 | 0 | — |
case-09 | fail→fail | 24,992 | 7,869 | -69% | 1 | 1 | 0% | 174 | 719 | +313% | 0 | 0 | — |
case-10 | fail→fail | 15,467 | 13,966 | -10% | 1 | 1 | 0% | 181 | 870 | +381% | 0 | 0 | — |
case-11 | fail→fail | 28,797 | 9,638 | -67% | 1 | 1 | 0% | 2,120 | 1,092 | -48% | 0 | 0 | — |
case-12 | fail→fail | 12,499 | 8,727 | -30% | 1 | 1 | 0% | 286 | 703 | +146% | 0 | 0 | — |
case-13 | fail→fail | 8,074 | 9,090 | +13% | 1 | 1 | 0% | 262 | 731 | +179% | 0 | 0 | — |
case-14 | fail→fail | 12,040 | 7,183 | -40% | 1 | 1 | 0% | 1,963 | 774 | -61% | 0 | 0 | — |
case-15 | fail→fail | 38,734 | 7,275 | -81% | 1 | 1 | 0% | 2,850 | 694 | -76% | 0 | 0 | — |
case-16 | fail→fail | 7,220 | 8,951 | +24% | 1 | 1 | 0% | 218 | 945 | +333% | 0 | 0 | — |
case-17 | pass→pass | 11,748 | 7,518 | -36% | 1 | 1 | 0% | 1,532 | 1,531 | -0% | 0 | 0 | — |
case-18 | fail→fail | 9,679 | 9,426 | -3% | 1 | 1 | 0% | 287 | 800 | +179% | 0 | 0 | — |
case-19 | fail→fail | 11,820 | 6,189 | -48% | 1 | 1 | 0% | 1,807 | 663 | -63% | 0 | 0 | — |
case-20 | pass→fail | 22,206 | 11,444 | -48% | 1 | 1 | 0% | 3,421 | 2,284 | -33% | 0 | 0 | — |
case-21 | pass→fail | 8,627 | 6,999 | -19% | 1 | 1 | 0% | 1,495 | 703 | -53% | 0 | 0 | — |
case-22 | pass→fail | 12,727 | 21,962 | +73% | 1 | 1 | 0% | 2,259 | 3,929 | +74% | 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. 22 cases were attempted, and 4 counted toward the lift figure. The other 18 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of -33 percentage points is the difference between those two pass rates over the 4 comparable cases. 5 cases got worse with the skill loaded, and they are 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.