Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when `evidence-review` needs an operational protocol before screening and extraction. **Trigger**: protocol, PRISMA, systematic review, inclusion/exclusion, 检索式, 纳入排除. **Use when**: `evidence-review` pipeline 的起点（C1），需要先锁定 protocol 再开始 screening/extraction。 **Skip if**: 不是做 evidence/systematic review（或 protocol 已经锁定且不允许修改）。 **Network**: none. **Guardrail**: protocol 必须包含可执行的检索与筛选规则；需要 HUMAN 签字后才能进入 screening。
.claude/skills/willoscar-protocol-writer/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | -66% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -76% | 0% |
| case-22 | ✗→✓ | ▲ Improved | -39% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -59% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -49% | 0% |
Transforms the review question into an executable evidence-review protocol.
Required:
STATUS.mdOptional:
GOAL.mdDECISIONS.mdqueries.mdoutput/PROTOCOL.mdThe protocol must contain:
scripts/run.py should:
It should not perform retrieval or screening itself.
output/PROTOCOL.md exists| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 20,518 | 3,592 | -82% | 1 | 1 | 0% | 3,554 | 400 | -89% | 0 | 0 | — |
case-02 | fail→fail | 18,291 | 3,008 | -84% | 1 | 1 | 0% | 3,002 | 342 | -89% | 0 | 0 | — |
case-03 | fail→fail | 17,959 | 3,796 | -79% | 1 | 1 | 0% | 2,857 | 391 | -86% | 0 | 0 | — |
case-04 | fail→pass | 8,799 | 1,699 | -81% | 1 | 1 | 0% | 1,400 | 482 | -66% | 0 | 0 | — |
case-05 | fail→pass | 12,095 | 1,843 | -85% | 1 | 1 | 0% | 1,731 | 422 | -76% | 0 | 0 | — |
case-22 | fail→pass | 11,467 | 5,873 | -49% | 1 | 1 | 0% | 1,710 | 1,035 | -39% | 0 | 0 | — |
case-06 | pass→pass | 15,102 | 3,921 | -74% | 1 | 1 | 0% | 2,234 | 765 | -66% | 0 | 0 | — |
case-07 | fail→pass | 13,908 | 4,362 | -69% | 1 | 1 | 0% | 2,041 | 837 | -59% | 0 | 0 | — |
case-08 | fail→pass | 13,041 | 5,142 | -61% | 1 | 1 | 0% | 2,070 | 1,048 | -49% | 0 | 0 | — |
case-09 | pass→pass | 17,240 | 8,617 | -50% | 1 | 1 | 0% | 2,835 | 1,508 | -47% | 0 | 0 | — |
case-10 | pass→fail | 9,538 | 2,968 | -69% | 1 | 1 | 0% | 1,530 | 691 | -55% | 0 | 0 | — |
case-11 | fail→pass | 5,380 | 3,981 | -26% | 1 | 1 | 0% | 836 | 807 | -3% | 0 | 0 | — |
case-12 | pass→fail | 5,580 | 3,746 | -33% | 1 | 1 | 0% | 830 | 682 | -18% | 0 | 0 | — |
case-13 | fail→pass | 5,067 | 11,983 | +136% | 1 | 1 | 0% | 771 | 2,197 | +185% | 0 | 0 | — |
case-14 | pass→pass | 21,916 | 4,856 | -78% | 1 | 1 | 0% | 3,174 | 969 | -69% | 0 | 0 | — |
case-15 | pass→pass | 7,685 | 18,559 | +141% | 1 | 1 | 0% | 1,160 | 3,327 | +187% | 0 | 0 | — |
case-16 | fail→pass | 8,053 | 2,342 | -71% | 1 | 1 | 0% | 1,403 | 544 | -61% | 0 | 0 | — |
case-17 | pass→pass | 14,756 | 11,274 | -24% | 1 | 1 | 0% | 2,167 | 1,905 | -12% | 0 | 0 | — |
case-18 | pass→pass | 6,785 | 3,957 | -42% | 1 | 1 | 0% | 981 | 831 | -15% | 0 | 0 | — |
case-19 | pass→pass | 10,654 | 3,550 | -67% | 1 | 1 | 0% | 1,537 | 765 | -50% | 0 | 0 | — |
case-20 | fail→pass | 6,931 | 3,155 | -54% | 1 | 1 | 0% | 995 | 727 | -27% | 0 | 0 | — |
case-21 | pass→pass | 11,088 | 5,298 | -52% | 1 | 1 | 0% | 1,734 | 1,027 | -41% | 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 19 counted toward the lift figure. The other 3 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 +32 percentage points is the difference between those two pass rates over the 19 comparable cases. 2 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.