Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Discover, evaluate, safely install, and update reusable skills from a curated set of trusted open-source repositories.
.claude/skills/sonichi-trusted-capabilities/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | -42% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -36% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -11% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -59% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -39% | 0% |
Use this skill when the owner asks to find, assess, install, or update reusable agent skills and tools from public ecosystems.
The catalog is allowlist-only. Skill sources can be installed after inspection. Tool sources can be discovered and statically inspected, while tool and index sources remain install-disabled because their setup and runtime permissions are project-specific.
bashC=skills/trusted-capabilities/scripts/catalog.py python3 "$C" sources python3 "$C" search browser python3 "$C" inspect openai-skills skills/.curated/browser-automation python3 "$C" inspect mcp-reference-servers src/filesystem python3 "$C" install openai-skills skills/.curated/browser-automation # Review the dry-run output, then copy its exact commit into the write: python3 "$C" install openai-skills skills/.curated/browser-automation --commit <40-char-sha> --yes python3 "$C" update browser-automation python3 "$C" update browser-automation --commit <40-char-sha> --yes
install and update show the resolved commit, destination, and risk findings, then print a write command pinned to that immutable commit. Writes require both the full --commit <sha> from the reviewed dry run and --yes; resolving the moving default branch again is not authorization to write. Updates also re-check that the source remains installable in the current allowlist. The default destination is the runtime's canonical $CLAUDE_CONFIG_DIR/skills/ path via src/util_paths.py. Use --dest-root for an explicit test or alternate install root.
Every install is pinned to a Git commit, fetched file-by-file, bounded by file count and total size, and written atomically. Provenance is recorded in .sutando-source.json; update uses it to compare and fetch the same upstream path.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-04 | fail→pass | 10,787 | 4,299 | -60% | 1 | 1 | 0% | 1,812 | 1,056 | -42% | 0 | 0 | — |
case-01 | fail→fail | 8,320 | 13,546 | +63% | 1 | 1 | 0% | 1,170 | 646 | -45% | 0 | 0 | — |
case-02 | fail→fail | 10,783 | 4,853 | -55% | 1 | 1 | 0% | 1,615 | 606 | -62% | 0 | 0 | — |
case-03 | fail→fail | 14,736 | 4,006 | -73% | 1 | 1 | 0% | 2,218 | 594 | -73% | 0 | 0 | — |
case-05 | fail→pass | 10,205 | 3,262 | -68% | 1 | 1 | 0% | 1,534 | 976 | -36% | 0 | 0 | — |
case-06 | fail→pass | 9,541 | 5,205 | -45% | 1 | 1 | 0% | 1,365 | 1,212 | -11% | 0 | 0 | — |
case-07 | fail→pass | 14,694 | 3,003 | -80% | 1 | 1 | 0% | 2,200 | 896 | -59% | 0 | 0 | — |
case-08 | fail→pass | 9,132 | 2,927 | -68% | 1 | 1 | 0% | 1,466 | 901 | -39% | 0 | 0 | — |
case-13 | fail→pass | 15,471 | 14,132 | -9% | 1 | 1 | 0% | 2,412 | 2,253 | -7% | 0 | 0 | — |
case-09 | fail→pass | 7,856 | 1,810 | -77% | 1 | 1 | 0% | 1,239 | 682 | -45% | 0 | 0 | — |
case-10 | fail→pass | 13,887 | 1,833 | -87% | 1 | 1 | 0% | 2,255 | 730 | -68% | 0 | 0 | — |
case-11 | pass→pass | 12,106 | 4,157 | -66% | 1 | 1 | 0% | 1,757 | 1,001 | -43% | 0 | 0 | — |
case-12 | fail→pass | 9,034 | 3,460 | -62% | 1 | 1 | 0% | 1,390 | 945 | -32% | 0 | 0 | — |
case-14 | fail→pass | 13,044 | 4,198 | -68% | 1 | 1 | 0% | 2,116 | 804 | -62% | 0 | 0 | — |
case-15 | fail→pass | 8,191 | 2,413 | -71% | 1 | 1 | 0% | 1,405 | 906 | -36% | 0 | 0 | — |
case-16 | pass→pass | 2,393 | 2,777 | +16% | 1 | 1 | 0% | 430 | 839 | +95% | 0 | 0 | — |
case-17 | pass→pass | 4,801 | 3,763 | -22% | 1 | 1 | 0% | 675 | 952 | +41% | 0 | 0 | — |
case-18 | pass→pass | 3,500 | 3,191 | -9% | 1 | 1 | 0% | 552 | 978 | +77% | 0 | 0 | — |
case-19 | fail→pass | 10,413 | 2,333 | -78% | 1 | 1 | 0% | 1,505 | 808 | -46% | 0 | 0 | — |
case-20 | fail→pass | 7,938 | 2,835 | -64% | 1 | 1 | 0% | 1,310 | 880 | -33% | 0 | 0 | — |
case-21 | fail→pass | 7,889 | 5,837 | -26% | 1 | 1 | 0% | 1,359 | 1,230 | -9% | 0 | 0 | — |
case-22 | fail→pass | 4,515 | 2,657 | -41% | 1 | 1 | 0% | 794 | 816 | +3% | 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 19 counted toward the lift figure. The other 3 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 +68 percentage points is the difference between those two pass rates over the 19 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.