Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when `paper-review` has claims plus evidence gaps and needs the final referee-style report. **Trigger**: rubric review, referee report, peer review write-up, 审稿报告, REVIEW.md. **Use when**: `paper-review` pipeline 的最后阶段（C3），已有 `output/CLAIMS.md` + `output/MISSING_EVIDENCE.md`（以及可选 novelty matrix）。
.claude/skills/willoscar-rubric-writer/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-09 | ✗→✓ | ▲ Improved | -34% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -75% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -25% | 0% |
| case-14 | ✗→✓ | ▲ Improved | 56% | 0% |
| case-20 | ✗→✓ | ▲ Improved | -80% | 0% |
Transforms review evidence artifacts into the final paper-review deliverable.
Required:
output/CLAIMS.mdoutput/MISSING_EVIDENCE.mdOptional:
output/NOVELTY_MATRIX.mdDECISIONS.mdoutput/REVIEW.mdThe review must expose stable sections:
### Summary### Novelty### Soundness### Clarity### Impact### Major Concerns### Minor Comments### RecommendationEvery major concern must cite its claim_id or gap_id. This is the traceability interface consumed by the paper-review scorecard.
scripts/run.py should:
It should not re-parse the manuscript from scratch or perform retrieval.
output/REVIEW.md exists| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | pass→fail | 12,580 | 16,541 | +31% | 1 | 1 | 0% | 2,348 | 3,114 | +33% | 0 | 0 | — |
case-02 | fail→fail | 4,517 | 4,714 | +4% | 1 | 1 | 0% | 232 | 445 | +92% | 0 | 0 | — |
case-03 | pass→pass | 38,643 | 10,531 | -73% | 1 | 1 | 0% | 4,961 | 2,020 | -59% | 0 | 0 | — |
case-04 | fail→fail | 4,394 | 3,802 | -13% | 1 | 1 | 0% | 742 | 406 | -45% | 0 | 0 | — |
case-05 | pass→fail | 17,034 | 3,815 | -78% | 1 | 1 | 0% | 2,741 | 406 | -85% | 0 | 0 | — |
case-06 | pass→fail | 15,417 | 5,948 | -61% | 1 | 1 | 0% | 2,391 | 474 | -80% | 0 | 0 | — |
case-07 | pass→pass | 16,368 | 6,231 | -62% | 1 | 1 | 0% | 2,528 | 1,362 | -46% | 0 | 0 | — |
case-08 | fail→fail | 9,221 | 2,617 | -72% | 1 | 1 | 0% | 1,387 | 686 | -51% | 0 | 0 | — |
case-09 | fail→pass | 9,327 | 3,117 | -67% | 1 | 1 | 0% | 1,362 | 900 | -34% | 0 | 0 | — |
case-10 | pass→pass | 13,195 | 3,482 | -74% | 1 | 1 | 0% | 1,869 | 889 | -52% | 0 | 0 | — |
case-11 | fail→pass | 27,210 | 1,589 | -94% | 1 | 1 | 0% | 1,958 | 481 | -75% | 0 | 0 | — |
case-12 | fail→pass | 6,604 | 3,569 | -46% | 1 | 1 | 0% | 999 | 753 | -25% | 0 | 0 | — |
case-13 | fail→fail | 9,458 | 4,001 | -58% | 1 | 1 | 0% | 1,413 | 919 | -35% | 0 | 0 | — |
case-14 | fail→pass | 10,178 | 13,697 | +35% | 1 | 1 | 0% | 1,513 | 2,367 | +56% | 0 | 0 | — |
case-15 | fail→fail | 7,583 | 3,247 | -57% | 1 | 1 | 0% | 1,169 | 731 | -37% | 0 | 0 | — |
case-16 | fail→fail | 11,050 | 2,257 | -80% | 1 | 1 | 0% | 1,660 | 609 | -63% | 0 | 0 | — |
case-17 | pass→pass | 4,736 | 2,128 | -55% | 1 | 1 | 0% | 873 | 565 | -35% | 0 | 0 | — |
case-18 | fail→fail | 10,426 | 5,743 | -45% | 1 | 1 | 0% | 1,835 | 1,154 | -37% | 0 | 0 | — |
case-19 | pass→pass | 11,301 | 2,844 | -75% | 1 | 1 | 0% | 1,737 | 728 | -58% | 0 | 0 | — |
case-20 | fail→pass | 11,991 | 1,530 | -87% | 1 | 1 | 0% | 2,325 | 475 | -80% | 0 | 0 | — |
case-21 | fail→fail | 9,227 | 8,948 | -3% | 1 | 1 | 0% | 1,610 | 2,262 | +40% | 0 | 0 | — |
case-22 | pass→fail | 19,618 | 8,523 | -57% | 1 | 1 | 0% | 3,040 | 1,534 | -50% | 0 | 0 | — |
case-23 | pass→pass | 13,607 | 11,699 | -14% | 1 | 1 | 0% | 2,015 | 1,967 | -2% | 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. 23 cases were attempted, and 17 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 +4 percentage points is the difference between those two pass rates over the 17 comparable cases. 4 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.