Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when an approved `evidence-review` protocol needs to be applied to a candidate pool. **Trigger**: screening, title/abstract screening, inclusion/exclusion, screening_log.csv, 文献筛选, 纳入排除. **Use when**: `evidence-review` 的 screening 阶段（C2/C3），protocol 已锁定并通过 HUMAN 审批。 **Skip if**: 还没有 `output/PROTOCOL.md`（或 protocol 未通过签字）。 **Network**: none. **Guardrail**: 每条记录包含决策与理由；保持可审计（不要把“未读/不确定”当作纳入）。
.claude/skills/willoscar-screening-manager/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-21 | ✗→✓ | ▲ Improved | -47% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -27% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -80% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 10% | 0% |
| case-15 | ✗→✓ | ▲ Improved | -70% | 0% |
Transforms an approved protocol plus candidate pool into an auditable screening log.
Required:
output/PROTOCOL.mdCandidate pool:
papers/papers_raw.jsonlpapers/papers_dedup.jsonlpapers/core_set.csvpapers/screening_log.csvEach row must include at least:
paper_idtitleyearurldecisionreasonreason_codesreviewerdecided_atscripts/run.py should:
It should not invent new protocol rules.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-21 | fail→pass | 14,567 | 4,548 | -69% | 1 | 1 | 0% | 2,092 | 1,102 | -47% | 0 | 0 | — |
case-01 | fail→fail | 5,817 | 3,888 | -33% | 1 | 1 | 0% | 361 | 458 | +27% | 0 | 0 | — |
case-02 | fail→fail | 5,954 | 33,925 | +470% | 1 | 1 | 0% | 427 | 520 | +22% | 0 | 0 | — |
case-03 | fail→fail | 9,067 | 4,319 | -52% | 1 | 1 | 0% | 235 | 375 | +60% | 0 | 0 | — |
case-04 | pass→pass | 14,587 | 5,719 | -61% | 1 | 1 | 0% | 2,606 | 1,169 | -55% | 0 | 0 | — |
case-09 | fail→pass | 8,220 | 4,431 | -46% | 1 | 1 | 0% | 1,289 | 947 | -27% | 0 | 0 | — |
case-05 | fail→pass | 40,397 | 6,378 | -84% | 1 | 1 | 0% | 6,172 | 1,226 | -80% | 0 | 0 | — |
case-06 | fail→fail | 23,134 | 21,474 | -7% | 1 | 1 | 0% | 3,704 | 3,804 | +3% | 0 | 0 | — |
case-07 | fail→fail | 3,831 | 4,531 | +18% | 1 | 1 | 0% | 204 | 369 | +81% | 0 | 0 | — |
case-08 | pass→fail | 10,810 | 7,595 | -30% | 1 | 1 | 0% | 1,854 | 1,483 | -20% | 0 | 0 | — |
case-19 | fail→fail | 21,452 | 4,384 | -80% | 1 | 1 | 0% | 193 | 402 | +108% | 0 | 0 | — |
case-10 | fail→pass | 18,237 | 17,486 | -4% | 1 | 1 | 0% | 3,160 | 3,462 | +10% | 0 | 0 | — |
case-11 | fail→fail | 18,309 | 2,028 | -89% | 1 | 1 | 0% | 1,196 | 543 | -55% | 0 | 0 | — |
case-12 | fail→fail | 13,401 | 1,729 | -87% | 1 | 1 | 0% | 1,957 | 441 | -77% | 0 | 0 | — |
case-13 | fail→fail | 8,604 | 1,860 | -78% | 1 | 1 | 0% | 1,319 | 520 | -61% | 0 | 0 | — |
case-20 | fail→fail | 6,474 | 3,157 | -51% | 1 | 1 | 0% | 959 | 689 | -28% | 0 | 0 | — |
case-14 | fail→fail | 9,644 | 3,892 | -60% | 1 | 1 | 0% | 1,498 | 859 | -43% | 0 | 0 | — |
case-15 | fail→pass | 10,546 | 2,170 | -79% | 1 | 1 | 0% | 1,772 | 540 | -70% | 0 | 0 | — |
case-16 | pass→pass | 3,711 | 2,861 | -23% | 1 | 1 | 0% | 575 | 739 | +29% | 0 | 0 | — |
case-17 | pass→pass | 8,165 | 2,923 | -64% | 1 | 1 | 0% | 1,243 | 689 | -45% | 0 | 0 | — |
case-18 | pass→pass | 5,489 | 2,506 | -54% | 1 | 1 | 0% | 800 | 647 | -19% | 0 | 0 | — |
case-22 | pass→fail | 9,147 | 3,985 | -56% | 1 | 1 | 0% | 1,352 | 877 | -35% | 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 16 counted toward the lift figure. The other 6 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 +14 percentage points is the difference between those two pass rates over the 16 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.