Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Lightweight anti-hallucination workflow for task kickoff, review prioritization, and regression retrospectives. Use when the user asks for guardrails, task contracts, risk scoring, or review templates.
.claude/skills/majiayu000-vibeguard/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-14 | ✗→✓ | ▲ Improved | -26% | 0% |
| case-01 | ✗→✓ | ▲ Improved | 129% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -49% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 47% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -62% | 0% |
Lightweight VibeGuard for everyday use inside Spellbook.
This skill helps with three moments where AI-assisted work usually drifts:
This skill is intentionally lightweight.
It does:
It does not:
If the user asks for automated interception, repo-level rules, or environment setup, escalate to the full VibeGuard repository and tooling.
Trigger this skill when the user asks for:
Open references/task-contract.yaml and confirm:
Do not move into implementation until these are concrete.
Open references/scoring-matrix.md and score each finding on:
Use the score to separate urgent fixes from weak guesses.
If something regressed, identify which defense failed:
Then capture the follow-up in references/review-template.md.
Depending on the request, produce one of these:
references/task-contract.yaml - kickoff checklistreferences/scoring-matrix.md - prioritization modelreferences/review-template.md - retrospective template| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-14 | fail→pass | 15,000 | 6,029 | -60% | 1 | 1 | 0% | 2,076 | 1,535 | -26% | 0 | 0 | — |
case-01 | fail→pass | 22,175 | 7,697 | -65% | 1 | 1 | 0% | 777 | 1,781 | +129% | 0 | 0 | — |
case-02 | pass→pass | 15,930 | 14,174 | -11% | 1 | 1 | 0% | 2,525 | 2,293 | -9% | 0 | 0 | — |
case-03 | fail→pass | 13,714 | 5,264 | -62% | 1 | 1 | 0% | 2,634 | 1,344 | -49% | 0 | 0 | — |
case-04 | fail→pass | 6,916 | 4,473 | -35% | 1 | 1 | 0% | 879 | 1,288 | +47% | 0 | 0 | — |
case-05 | fail→pass | 17,969 | 3,981 | -78% | 1 | 1 | 0% | 3,084 | 1,178 | -62% | 0 | 0 | — |
case-06 | fail→pass | 15,623 | 8,924 | -43% | 1 | 1 | 0% | 2,396 | 2,031 | -15% | 0 | 0 | — |
case-07 | pass→pass | 15,841 | 8,155 | -49% | 1 | 1 | 0% | 2,510 | 1,715 | -32% | 0 | 0 | — |
case-08 | pass→pass | 10,910 | 3,291 | -70% | 1 | 1 | 0% | 1,598 | 1,134 | -29% | 0 | 0 | — |
case-09 | fail→pass | 14,693 | 4,703 | -68% | 1 | 1 | 0% | 2,299 | 1,376 | -40% | 0 | 0 | — |
case-10 | pass→pass | 16,479 | 7,229 | -56% | 1 | 1 | 0% | 2,317 | 1,659 | -28% | 0 | 0 | — |
case-11 | pass→pass | 7,760 | 4,078 | -47% | 1 | 1 | 0% | 1,109 | 1,219 | +10% | 0 | 0 | — |
case-12 | fail→pass | 8,745 | 3,342 | -62% | 1 | 1 | 0% | 1,247 | 1,084 | -13% | 0 | 0 | — |
case-13 | fail→pass | 17,081 | 13,540 | -21% | 1 | 1 | 0% | 2,767 | 2,838 | +3% | 0 | 0 | — |
case-15 | fail→pass | 16,948 | 10,043 | -41% | 1 | 1 | 0% | 2,551 | 2,156 | -15% | 0 | 0 | — |
case-16 | pass→pass | 8,059 | 3,053 | -62% | 1 | 1 | 0% | 1,261 | 1,046 | -17% | 0 | 0 | — |
case-17 | fail→pass | 14,474 | 10,462 | -28% | 1 | 1 | 0% | 2,074 | 2,264 | +9% | 0 | 0 | — |
case-18 | fail→pass | 11,218 | 6,104 | -46% | 1 | 1 | 0% | 1,641 | 1,534 | -7% | 0 | 0 | — |
case-19 | fail→pass | 11,128 | 5,697 | -49% | 1 | 1 | 0% | 1,643 | 1,374 | -16% | 0 | 0 | — |
case-20 | pass→pass | 11,824 | 6,076 | -49% | 1 | 1 | 0% | 1,714 | 1,550 | -10% | 0 | 0 | — |
case-21 | pass→pass | 8,870 | 4,607 | -48% | 1 | 1 | 0% | 1,293 | 1,300 | +1% | 0 | 0 | — |
case-22 | fail→pass | 13,925 | 2,644 | -81% | 1 | 1 | 0% | 2,182 | 1,010 | -54% | 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. The headline lift of +64 percentage points is the difference between those two pass rates over the 22 comparable cases.
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.