Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Orchestrate pi-finder and optional pi-librarian to produce a compact context pack and Codex kickoff prompt before coding.
.claude/skills/cowork-os-pi-context-pipeline/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-03 | ✗→✓ | ▲ Improved | 63% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 75% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -36% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -6% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 66% | 0% |
Orchestrate pi-finder and optional pi-librarian to produce a compact context pack and Codex kickoff prompt before coding.
| Name | Type | Required | Description | |---|---|---|---| | query | string | Yes | What to implement or investigate | | scope_hint | string | No | Optional local directories/files/symbols to prioritize | | repo_hints | string | No | Optional external repo hints (owner/repo list) | | owner_hints | string | No | Optional org/user hints for GitHub research | | run_librarian | boolean | No | Whether to include GitHub research in addition to local scouting | | max_search_results | number | No | Maximum GitHub search results requested by librarian |
../pi-context-pipeline.json.| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 10,122 | 7,577 | -25% | 1 | 1 | 0% | 1,656 | 1,700 | +3% | 0 | 0 | — |
case-02 | fail→fail | 12,155 | 9,275 | -24% | 1 | 1 | 0% | 2,173 | 2,134 | -2% | 0 | 0 | — |
case-03 | fail→pass | 13,350 | 22,091 | +65% | 1 | 1 | 0% | 2,143 | 3,497 | +63% | 0 | 0 | — |
case-04 | fail→pass | 3,411 | 2,835 | -17% | 1 | 1 | 0% | 550 | 961 | +75% | 0 | 0 | — |
case-05 | pass→pass | 4,411 | 2,446 | -45% | 1 | 1 | 0% | 692 | 800 | +16% | 0 | 0 | — |
case-06 | pass→pass | 3,674 | 3,367 | -8% | 1 | 1 | 0% | 595 | 979 | +65% | 0 | 0 | — |
case-07 | fail→pass | 11,412 | 5,102 | -55% | 1 | 1 | 0% | 1,938 | 1,247 | -36% | 0 | 0 | — |
case-08 | pass→fail | 17,725 | 6,373 | -64% | 1 | 1 | 0% | 3,214 | 1,502 | -53% | 0 | 0 | — |
case-09 | fail→pass | 6,285 | 3,036 | -52% | 1 | 1 | 0% | 1,054 | 992 | -6% | 0 | 0 | — |
case-10 | fail→pass | 8,527 | 13,302 | +56% | 1 | 1 | 0% | 1,252 | 2,082 | +66% | 0 | 0 | — |
case-11 | fail→fail | 9,877 | 4,304 | -56% | 1 | 1 | 0% | 1,762 | 1,178 | -33% | 0 | 0 | — |
case-12 | pass→pass | 9,336 | 5,222 | -44% | 1 | 1 | 0% | 1,398 | 1,300 | -7% | 0 | 0 | — |
case-13 | fail→pass | 10,552 | 2,188 | -79% | 1 | 1 | 0% | 1,688 | 884 | -48% | 0 | 0 | — |
case-14 | fail→pass | 9,055 | 4,306 | -52% | 1 | 1 | 0% | 1,420 | 1,131 | -20% | 0 | 0 | — |
case-15 | fail→pass | 11,407 | 5,309 | -53% | 1 | 1 | 0% | 1,799 | 1,347 | -25% | 0 | 0 | — |
case-16 | fail→fail | 5,756 | 2,070 | -64% | 1 | 1 | 0% | 853 | 799 | -6% | 0 | 0 | — |
case-17 | fail→fail | 11,555 | 2,107 | -82% | 1 | 1 | 0% | 1,752 | 752 | -57% | 0 | 0 | — |
case-18 | pass→pass | 9,537 | 3,031 | -68% | 1 | 1 | 0% | 1,384 | 975 | -30% | 0 | 0 | — |
case-19 | pass→pass | 9,444 | 3,604 | -62% | 1 | 1 | 0% | 1,507 | 986 | -35% | 0 | 0 | — |
case-20 | fail→pass | 11,967 | 6,342 | -47% | 1 | 1 | 0% | 1,871 | 1,447 | -23% | 0 | 0 | — |
case-21 | fail→pass | 12,652 | 4,003 | -68% | 1 | 1 | 0% | 1,795 | 1,069 | -40% | 0 | 0 | — |
case-22 | pass→pass | 10,551 | 2,969 | -72% | 1 | 1 | 0% | 1,600 | 893 | -44% | 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 +41 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.