Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when a tutorial module plan exists and the run needs an auditable module-to-source grounding file before prose. **Trigger**: module coverage, source coverage, tutorial grounding, 模块覆盖, 来源覆盖. **Use when**: `source-tutorial` 的 C2，已有 `outline/module_plan.yml`，需要确认每个模块都能回指到 sources。 **Skip if**: module plan 或 source ingest 不完整。 **Network**: none. **Guardrail**: 只做 grounding audit，不写教程正文。
.claude/skills/willoscar-module-source-coverage/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 130% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 48% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -13% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -41% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -64% | 0% |
Builds outline/source_coverage.jsonl, one coverage record per module.
outline/module_plan.ymlsources/index.jsonlsources/provenance.jsonloutline/source_coverage.jsonlEach record must include:
module_idmodule_titlesource_ids and/or explicit gapsscripts/run.py should:
Keep text matching and snippet scoring in shared tutorial tooling, not in the wrapper.
outline/source_coverage.jsonl exists| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 4,053 | 3,793 | -6% | 1 | 1 | 0% | 245 | 418 | +71% | 0 | 0 | — |
case-02 | fail→fail | 4,229 | 3,524 | -17% | 1 | 1 | 0% | 678 | 389 | -43% | 0 | 0 | — |
case-03 | fail→fail | 5,083 | 3,551 | -30% | 1 | 1 | 0% | 305 | 353 | +16% | 0 | 0 | — |
case-04 | fail→pass | 35,180 | 15,113 | -57% | 1 | 1 | 0% | 1,302 | 2,998 | +130% | 0 | 0 | — |
case-05 | pass→fail | 19,195 | 12,700 | -34% | 1 | 1 | 0% | 3,273 | 2,239 | -32% | 0 | 0 | — |
case-06 | fail→pass | 37,760 | 12,157 | -68% | 1 | 1 | 0% | 1,719 | 2,543 | +48% | 0 | 0 | — |
case-11 | fail→fail | 7,728 | 1,991 | -74% | 1 | 1 | 0% | 1,220 | 517 | -58% | 0 | 0 | — |
case-07 | fail→pass | 11,061 | 7,504 | -32% | 1 | 1 | 0% | 1,750 | 1,516 | -13% | 0 | 0 | — |
case-08 | fail→pass | 6,374 | 2,202 | -65% | 1 | 1 | 0% | 984 | 584 | -41% | 0 | 0 | — |
case-09 | fail→pass | 10,407 | 2,433 | -77% | 1 | 1 | 0% | 1,580 | 572 | -64% | 0 | 0 | — |
case-10 | fail→pass | 9,705 | 2,119 | -78% | 1 | 1 | 0% | 1,640 | 526 | -68% | 0 | 0 | — |
case-12 | fail→pass | 9,738 | 3,298 | -66% | 1 | 1 | 0% | 1,508 | 764 | -49% | 0 | 0 | — |
case-13 | fail→fail | 6,012 | 2,572 | -57% | 1 | 1 | 0% | 922 | 659 | -29% | 0 | 0 | — |
case-14 | pass→pass | 8,129 | 3,772 | -54% | 1 | 1 | 0% | 1,231 | 831 | -32% | 0 | 0 | — |
case-15 | fail→pass | 13,145 | 2,089 | -84% | 1 | 1 | 0% | 1,831 | 515 | -72% | 0 | 0 | — |
case-16 | fail→pass | 11,219 | 1,633 | -85% | 1 | 1 | 0% | 1,710 | 443 | -74% | 0 | 0 | — |
case-17 | pass→pass | 10,228 | 2,823 | -72% | 1 | 1 | 0% | 1,629 | 629 | -61% | 0 | 0 | — |
case-18 | pass→pass | 8,806 | 3,995 | -55% | 1 | 1 | 0% | 1,282 | 926 | -28% | 0 | 0 | — |
case-19 | pass→pass | 14,180 | 5,592 | -61% | 1 | 1 | 0% | 2,338 | 1,126 | -52% | 0 | 0 | — |
case-20 | pass→pass | 12,481 | 8,038 | -36% | 1 | 1 | 0% | 1,825 | 1,411 | -23% | 0 | 0 | — |
case-21 | fail→pass | 10,960 | 3,278 | -70% | 1 | 1 | 0% | 1,780 | 761 | -57% | 0 | 0 | — |
case-22 | pass→pass | 12,527 | 3,643 | -71% | 1 | 1 | 0% | 1,886 | 768 | -59% | 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 17 counted toward the lift figure. The other 5 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 +41 percentage points is the difference between those two pass rates over the 17 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.