Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Build and maintain a workspace-local, Obsidian-friendly research vault using Karpathy's LLM Wiki pattern.
.claude/skills/cowork-os-llm-wiki/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-06 | ✗→✓ | ▲ Improved | -21% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -47% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -72% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -69% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -47% | 0% |
Create and maintain a persistent markdown research vault with immutable raw sources, linked concept/entity pages, maps of content, and a maintenance log.
{artifactDir}/wiki-manifest.md and {artifactDir}/wiki-summary.md.index.md, log.md, and inbox.md remain current.| Name | Type | Required | Description | |---|---|---|---| | objective | string | No | Topic, question, or research objective. If omitted, ask one short scoping question before durable ingest work. | | mode | string | No | auto, init, ingest, query, lint, or refresh | | path | string | No | Workspace-relative or absolute vault path | | obsidian | string | No | auto, on, or off |
../llm-wiki.json.research/wiki relative to the current workspace.scripts/wiki-graph-report.mjs.scripts/wiki-import.mjs.scripts/wiki-search.mjs.scripts/wiki-render.mjs for Marp slide decks and SVG charts.| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 37,575 | 13,409 | -64% | 1 | 1 | 0% | 6,195 | 2,240 | -64% | 0 | 0 | — |
case-02 | fail→fail | 32,261 | 5,489 | -83% | 1 | 1 | 0% | 6,183 | 613 | -90% | 0 | 0 | — |
case-03 | pass→pass | 5,623 | 4,221 | -25% | 1 | 1 | 0% | 779 | 990 | +27% | 0 | 0 | — |
case-04 | pass→pass | 14,211 | 12,063 | -15% | 1 | 1 | 0% | 2,096 | 2,360 | +13% | 0 | 0 | — |
case-05 | pass→pass | 6,568 | 4,327 | -34% | 1 | 1 | 0% | 1,053 | 1,070 | +2% | 0 | 0 | — |
case-06 | fail→pass | 6,483 | 3,051 | -53% | 1 | 1 | 0% | 1,014 | 806 | -21% | 0 | 0 | — |
case-07 | fail→pass | 9,194 | 2,301 | -75% | 1 | 1 | 0% | 1,427 | 750 | -47% | 0 | 0 | — |
case-08 | fail→pass | 16,798 | 2,724 | -84% | 1 | 1 | 0% | 2,790 | 792 | -72% | 0 | 0 | — |
case-09 | fail→pass | 13,220 | 1,741 | -87% | 1 | 1 | 0% | 1,974 | 614 | -69% | 0 | 0 | — |
case-10 | fail→fail | 13,032 | 1,778 | -86% | 1 | 1 | 0% | 1,892 | 639 | -66% | 0 | 0 | — |
case-11 | fail→fail | 10,360 | 1,691 | -84% | 1 | 1 | 0% | 1,655 | 583 | -65% | 0 | 0 | — |
case-12 | fail→pass | 7,770 | 2,415 | -69% | 1 | 1 | 0% | 1,236 | 661 | -47% | 0 | 0 | — |
case-13 | fail→pass | 8,418 | 3,030 | -64% | 1 | 1 | 0% | 1,344 | 849 | -37% | 0 | 0 | — |
case-14 | fail→pass | 13,413 | 2,163 | -84% | 1 | 1 | 0% | 934 | 633 | -32% | 0 | 0 | — |
case-15 | fail→pass | 24,211 | 2,051 | -92% | 1 | 1 | 0% | 1,252 | 635 | -49% | 0 | 0 | — |
case-16 | pass→pass | 7,909 | 3,333 | -58% | 1 | 1 | 0% | 1,234 | 901 | -27% | 0 | 0 | — |
case-17 | fail→pass | 10,727 | 3,981 | -63% | 1 | 1 | 0% | 1,756 | 908 | -48% | 0 | 0 | — |
case-18 | fail→pass | 15,756 | 2,250 | -86% | 1 | 1 | 0% | 2,392 | 671 | -72% | 0 | 0 | — |
case-19 | fail→pass | 20,051 | 8,845 | -56% | 1 | 1 | 0% | 919 | 1,747 | +90% | 0 | 0 | — |
case-20 | fail→pass | 14,553 | 14,592 | +0% | 1 | 1 | 0% | 2,489 | 2,351 | -6% | 0 | 0 | — |
case-21 | fail→pass | 16,183 | 12,236 | -24% | 1 | 1 | 0% | 2,687 | 2,376 | -12% | 0 | 0 | — |
case-22 | fail→pass | 29,124 | 14,066 | -52% | 1 | 1 | 0% | 6,162 | 3,536 | -43% | 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 +64 percentage points is the difference between those two pass rates over the 18 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.