Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Select the minimal set of 1-3 verbatim sentences from a candidate paper/abstract sufficient to entail or refute an atomic claim (SciFact's rationale-selection step). Use this after claim-writing has produced an atomic claim, as the evidence-gathering step before claim-label-prediction; an empty rationale set is a valid outcome, not an error.
.claude/skills/yogsoth-ai-rationale-selection/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-06 | ✗→✓ | ▲ Improved | -12% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -43% | 0% |
| case-04 | ✓→✓ | ＝ Same ✓ | -3% | 0% |
| case-05 | ✓→✓ | ＝ Same ✓ | -15% | 0% |
| case-08 | ✓→✓ | ＝ Same ✓ | -17% | 0% |
Selects minimal evidentiary sentence set for a claim — middle step of the SciFact 3-chain, added to close coverage-audit finding S7 (the original graph jumped straight from claim-writing to a three-way label judgment with no evidence-selection step, even though the tag table's own stated output anchor explicitly requires rationale sentences alongside the label).
Subagent — spawned via spawn-agent skill.
<!-- BEGIN available-tables (generated) -->
| SOP | When to use | | --- | --- | | spawn-agent | Spawn a customized CC subagent with full MCP tool access. |
<!-- END available-tables (generated) -->
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-22 | fail→fail | 17,003 | 12,283 | -28% | 1 | 1 | 0% | 2,276 | 1,420 | -38% | 0 | 0 | — |
case-01 | fail→fail | 12,288 | 20,054 | +63% | 1 | 1 | 0% | 1,333 | 2,957 | +122% | 0 | 0 | — |
case-02 | fail→fail | 7,561 | 24,612 | +226% | 1 | 1 | 0% | 397 | 1,033 | +160% | 0 | 0 | — |
case-03 | fail→fail | 11,520 | 15,165 | +32% | 1 | 1 | 0% | 1,202 | 876 | -27% | 0 | 0 | — |
case-04 | pass→pass | 14,523 | 12,950 | -11% | 1 | 1 | 0% | 1,479 | 1,431 | -3% | 0 | 0 | — |
case-05 | pass→pass | 11,194 | 9,715 | -13% | 1 | 1 | 0% | 1,000 | 847 | -15% | 0 | 0 | — |
case-06 | fail→pass | 9,487 | 7,839 | -17% | 1 | 1 | 0% | 731 | 643 | -12% | 0 | 0 | — |
case-07 | fail→pass | 36,976 | 8,113 | -78% | 1 | 1 | 0% | 1,145 | 649 | -43% | 0 | 0 | — |
case-08 | pass→pass | 10,093 | 9,097 | -10% | 1 | 1 | 0% | 865 | 719 | -17% | 0 | 0 | — |
case-09 | pass→pass | 7,912 | 8,227 | +4% | 1 | 1 | 0% | 529 | 666 | +26% | 0 | 0 | — |
case-10 | fail→fail | 7,882 | 20,429 | +159% | 1 | 1 | 0% | 462 | 724 | +57% | 0 | 0 | — |
case-11 | fail→fail | 11,227 | 21,658 | +93% | 1 | 1 | 0% | 1,102 | 3,213 | +192% | 0 | 0 | — |
case-12 | pass→pass | 10,333 | 7,865 | -24% | 1 | 1 | 0% | 930 | 628 | -32% | 0 | 0 | — |
case-13 | pass→pass | 16,345 | 12,035 | -26% | 1 | 1 | 0% | 1,982 | 1,228 | -38% | 0 | 0 | — |
case-14 | fail→fail | 12,486 | 29,074 | +133% | 1 | 1 | 0% | 1,168 | 2,229 | +91% | 0 | 0 | — |
case-15 | pass→pass | 20,352 | 9,878 | -51% | 1 | 1 | 0% | 2,242 | 873 | -61% | 0 | 0 | — |
case-16 | fail→fail | 16,110 | 21,835 | +36% | 1 | 1 | 0% | 1,810 | 838 | -54% | 0 | 0 | — |
case-17 | pass→pass | 10,125 | 8,030 | -21% | 1 | 1 | 0% | 762 | 644 | -15% | 0 | 0 | — |
case-18 | fail→fail | 17,883 | 23,593 | +32% | 1 | 1 | 0% | 2,461 | 1,039 | -58% | 0 | 0 | — |
case-19 | pass→pass | 9,507 | 7,902 | -17% | 1 | 1 | 0% | 765 | 632 | -17% | 0 | 0 | — |
case-20 | pass→pass | 8,097 | 8,063 | -0% | 1 | 1 | 0% | 602 | 684 | +14% | 0 | 0 | — |
case-21 | pass→pass | 13,275 | 14,200 | +7% | 1 | 1 | 0% | 1,208 | 1,866 | +54% | 0 | 0 | — |
case-23 | pass→pass | 14,725 | 12,692 | -14% | 1 | 1 | 0% | 1,723 | 1,502 | -13% | 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 +9 percentage points is the difference between those two pass rates over the 17 comparable cases.
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.