Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Break down full ARR (ACL Rolling Review) reviewer responses into structured rebuttal units. Use when input contains ARR reviewer fields (Paper Summary, Strengths, Weaknesses, Comments/Suggestions) and numeric scores (Confidence, Soundness, Excitement, Overall Assessment, Reproducibility). Splits weaknesses and comments/suggestions into granular response items while preserving original wording for quoted issues.
.claude/skills/runtsang-stage1-arr-breakdown/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-05 | ✗→✓ | ▲ Improved | -45% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -58% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -40% | 0% |
| case-15 | ✗→✓ | ▲ Improved | -51% | 0% |
| case-21 | ✗→✓ | ▲ Improved | 158% | 0% |
Apply the shared template first:
skills/stage1/template/SKILL.mdThen apply only the ARR-specific overrides below.
Extract exactly these five keys (numbers only, decimals allowed):
confidence <- ARR Confidencesoundness <- ARR Soundnessexcitement <- ARR Excitementassessment <- ARR Overall Assessment (number before = sign)reproducibility <- ARR Reproducibility (number before = sign)summary <- Paper Summarystrength <- Summary Of StrengthsSummary Of WeaknessesComments Suggestions And Typos (label as suggestion)# Stage1 ARR Breakdown## Scores must contain exactly the five keys above.2.5, 3.5). Extract only the numeric part before any = or description text.Ethical Concerns, Datasets, Software, Knowledge Of Paper, and certification fields are irrelevant — ignore them.Comments Suggestions And Typos to the suggestion section and split it into atomic issues just as weaknesses are split.| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-03 | fail→fail | 4,304 | 6,877 | +60% | 1 | 1 | 0% | 702 | 1,072 | +53% | 0 | 0 | — |
case-04 | pass→pass | 13,538 | 3,856 | -72% | 1 | 1 | 0% | 1,805 | 896 | -50% | 0 | 0 | — |
case-01 | fail→fail | 6,876 | 5,115 | -26% | 1 | 1 | 0% | 1,189 | 1,206 | +1% | 0 | 0 | — |
case-02 | fail→fail | 10,451 | 11,249 | +8% | 1 | 1 | 0% | 1,581 | 1,870 | +18% | 0 | 0 | — |
case-05 | fail→pass | 12,628 | 4,732 | -63% | 1 | 1 | 0% | 1,783 | 975 | -45% | 0 | 0 | — |
case-06 | fail→pass | 14,004 | 2,441 | -83% | 1 | 1 | 0% | 1,827 | 771 | -58% | 0 | 0 | — |
case-07 | fail→pass | 10,505 | 3,702 | -65% | 1 | 1 | 0% | 1,616 | 973 | -40% | 0 | 0 | — |
case-08 | pass→pass | 12,930 | 3,913 | -70% | 1 | 1 | 0% | 1,719 | 821 | -52% | 0 | 0 | — |
case-09 | pass→pass | 11,945 | 3,030 | -75% | 1 | 1 | 0% | 1,912 | 658 | -66% | 0 | 0 | — |
case-10 | pass→pass | 14,701 | 1,773 | -88% | 1 | 1 | 0% | 1,736 | 551 | -68% | 0 | 0 | — |
case-11 | pass→fail | 11,078 | 2,128 | -81% | 1 | 1 | 0% | 1,478 | 539 | -64% | 0 | 0 | — |
case-12 | pass→pass | 7,790 | 5,091 | -35% | 1 | 1 | 0% | 1,370 | 896 | -35% | 0 | 0 | — |
case-13 | pass→pass | 13,474 | 4,239 | -69% | 1 | 1 | 0% | 1,860 | 892 | -52% | 0 | 0 | — |
case-14 | pass→pass | 6,331 | 1,973 | -69% | 1 | 1 | 0% | 933 | 632 | -32% | 0 | 0 | — |
case-15 | fail→pass | 10,393 | 2,572 | -75% | 1 | 1 | 0% | 1,435 | 702 | -51% | 0 | 0 | — |
case-16 | fail→fail | 9,364 | 2,747 | -71% | 1 | 1 | 0% | 1,360 | 753 | -45% | 0 | 0 | — |
case-17 | fail→fail | 14,620 | 2,673 | -82% | 1 | 1 | 0% | 2,728 | 690 | -75% | 0 | 0 | — |
case-18 | pass→pass | 16,134 | 4,144 | -74% | 1 | 1 | 0% | 2,032 | 920 | -55% | 0 | 0 | — |
case-19 | fail→fail | 15,377 | 2,406 | -84% | 1 | 1 | 0% | 2,019 | 522 | -74% | 0 | 0 | — |
case-20 | fail→fail | 9,513 | 20,246 | +113% | 1 | 1 | 0% | 1,481 | 3,209 | +117% | 0 | 0 | — |
case-21 | fail→pass | 9,731 | 26,870 | +176% | 1 | 1 | 0% | 1,586 | 4,086 | +158% | 0 | 0 | — |
case-22 | pass→pass | 22,146 | 16,130 | -27% | 1 | 1 | 0% | 2,557 | 2,261 | -12% | 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 +18 percentage points is the difference between those two pass rates over the 22 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.