Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when a review workspace has manuscript text and needs a traceable claim ledger. **Trigger**: claims extractor, extract claims, contributions, assumptions, peer review, 审稿, 主张提取. **Use when**: 审稿/评审或 evidence audit，需要把主张列表落盘并可追溯到原文位置（section/page/quote）。 **Skip if**: 没有可用的稿件/全文（例如缺少 `output/PAPER.md` 或等价文本）。 **Network**: none. **Guardrail**: 每条 claim 必须带可定位的 source pointer；区分 empirical vs conceptual claims。
.claude/skills/willoscar-claims-extractor/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 36% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -49% | 0% |
| case-15 | ✗→✓ | ▲ Improved | -71% | 0% |
| case-18 | ✗→✓ | ▲ Improved | -38% | 0% |
| case-20 | ✗→✓ | ▲ Improved | -64% | 0% |
Transforms manuscript text into a traceable claim ledger for paper-review.
output/PAPER.mdoutput/CLAIMS.mdoutput/CLAIMS.jsonl (review-claim.v1, one addressable claim per line)Each claim block must include:
empirical or conceptualoutput/PAPER.mdThe Markdown and JSONL files are two views of the same claim records. Downstream skills should prefer JSONL and retain Markdown fallback for older Workspaces.
scripts/run.py should:
Keep parsing and ranking heuristics in shared review tooling, not in the skill script.
output/CLAIMS.md existsoutput/CLAIMS.jsonl exists and every record has a stable claim_id| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 3,545 | 3,671 | +4% | 1 | 1 | 0% | 201 | 505 | +151% | 0 | 0 | — |
case-02 | fail→fail | 4,520 | 5,137 | +14% | 1 | 1 | 0% | 260 | 499 | +92% | 0 | 0 | — |
case-03 | fail→fail | 4,222 | 4,211 | -0% | 1 | 1 | 0% | 233 | 466 | +100% | 0 | 0 | — |
case-04 | fail→pass | 4,984 | 4,489 | -10% | 1 | 1 | 0% | 743 | 1,009 | +36% | 0 | 0 | — |
case-05 | fail→fail | 7,103 | 4,692 | -34% | 1 | 1 | 0% | 1,105 | 481 | -56% | 0 | 0 | — |
case-06 | fail→fail | 15,341 | 10,115 | -34% | 1 | 1 | 0% | 2,214 | 1,572 | -29% | 0 | 0 | — |
case-07 | pass→pass | 11,471 | 2,375 | -79% | 1 | 1 | 0% | 1,832 | 592 | -68% | 0 | 0 | — |
case-08 | pass→pass | 11,368 | 3,085 | -73% | 1 | 1 | 0% | 1,702 | 768 | -55% | 0 | 0 | — |
case-09 | pass→pass | 14,404 | 4,648 | -68% | 1 | 1 | 0% | 2,029 | 980 | -52% | 0 | 0 | — |
case-10 | pass→pass | 6,956 | 2,774 | -60% | 1 | 1 | 0% | 1,038 | 749 | -28% | 0 | 0 | — |
case-11 | fail→fail | 4,181 | 2,831 | -32% | 1 | 1 | 0% | 667 | 744 | +12% | 0 | 0 | — |
case-12 | fail→pass | 14,851 | 4,708 | -68% | 1 | 1 | 0% | 2,089 | 1,055 | -49% | 0 | 0 | — |
case-13 | pass→pass | 10,752 | 3,102 | -71% | 1 | 1 | 0% | 1,605 | 797 | -50% | 0 | 0 | — |
case-14 | fail→fail | 7,731 | 2,430 | -69% | 1 | 1 | 0% | 1,233 | 595 | -52% | 0 | 0 | — |
case-15 | fail→pass | 9,386 | 1,176 | -87% | 1 | 1 | 0% | 1,458 | 423 | -71% | 0 | 0 | — |
case-16 | pass→pass | 9,414 | 2,782 | -70% | 1 | 1 | 0% | 1,506 | 716 | -52% | 0 | 0 | — |
case-17 | pass→fail | 10,608 | 2,860 | -73% | 1 | 1 | 0% | 1,477 | 660 | -55% | 0 | 0 | — |
case-18 | fail→pass | 9,987 | 4,517 | -55% | 1 | 1 | 0% | 1,439 | 898 | -38% | 0 | 0 | — |
case-19 | fail→fail | 7,334 | 3,705 | -49% | 1 | 1 | 0% | 1,226 | 911 | -26% | 0 | 0 | — |
case-20 | fail→pass | 9,359 | 1,445 | -85% | 1 | 1 | 0% | 1,328 | 482 | -64% | 0 | 0 | — |
case-21 | pass→pass | 4,892 | 3,296 | -33% | 1 | 1 | 0% | 688 | 804 | +17% | 0 | 0 | — |
case-22 | fail→pass | 9,646 | 2,544 | -74% | 1 | 1 | 0% | 1,347 | 666 | -51% | 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 18 counted toward the lift figure. The other 4 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 +23 percentage points is the difference between those two pass rates over the 18 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.