Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Default reference pipeline for the tune-collab taskKind — pick a direction, patch-edit the existing artifact, critique, hand off.
.claude/skills/nexu-io-od-tune-collab/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | -66% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -43% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -48% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -62% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -57% | 0% |
Spec §1 / §10.1 / §21.5 / §23.3.3: the canonical "iterate on an existing artifact" flow. Inherits parentArtifactId from the project metadata so the produced artifact's ArtifactManifest.parentArtifactId is set automatically — the artifact lineage chain stays intact across multi-turn tune cycles.
jsonc{ "stages": [ { "id": "direction", "atoms": ["direction-picker"] }, { "id": "patch", "atoms": ["patch-edit"] }, { "id": "critique", "atoms": ["critique-theater"], "repeat": true, "until": "critique.score>=4 || iterations>=3" }, { "id": "handoff", "atoms": ["handoff"] } ] }
The handoff stage records handoffKind: 'patch' (or 'deployable-app' when the user opted in via od plugin run --target deployable-app AND build-test ran successfully somewhere upstream in the project's history).
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-21 | pass→pass | 13,512 | 12,807 | -5% | 1 | 1 | 0% | 2,270 | 2,636 | +16% | 0 | 0 | — |
case-01 | fail→pass | 38,478 | 16,324 | -58% | 1 | 1 | 0% | 8,247 | 2,785 | -66% | 0 | 0 | — |
case-02 | fail→fail | 22,632 | 14,714 | -35% | 1 | 1 | 0% | 4,519 | 2,592 | -43% | 0 | 0 | — |
case-03 | fail→pass | 26,038 | 12,913 | -50% | 1 | 1 | 0% | 4,865 | 2,769 | -43% | 0 | 0 | — |
case-04 | fail→pass | 14,177 | 5,518 | -61% | 1 | 1 | 0% | 2,696 | 1,393 | -48% | 0 | 0 | — |
case-05 | fail→pass | 9,962 | 2,212 | -78% | 1 | 1 | 0% | 1,706 | 645 | -62% | 0 | 0 | — |
case-06 | fail→pass | 12,735 | 2,696 | -79% | 1 | 1 | 0% | 1,899 | 820 | -57% | 0 | 0 | — |
case-07 | pass→fail | 15,460 | 4,359 | -72% | 1 | 1 | 0% | 2,170 | 622 | -71% | 0 | 0 | — |
case-08 | fail→pass | 22,786 | 2,145 | -91% | 1 | 1 | 0% | 1,466 | 716 | -51% | 0 | 0 | — |
case-09 | fail→pass | 9,593 | 2,902 | -70% | 1 | 1 | 0% | 1,688 | 823 | -51% | 0 | 0 | — |
case-10 | fail→pass | 8,308 | 2,524 | -70% | 1 | 1 | 0% | 1,380 | 727 | -47% | 0 | 0 | — |
case-11 | fail→pass | 7,515 | 2,867 | -62% | 1 | 1 | 0% | 1,241 | 768 | -38% | 0 | 0 | — |
case-12 | pass→pass | 5,893 | 2,980 | -49% | 1 | 1 | 0% | 987 | 855 | -13% | 0 | 0 | — |
case-13 | fail→pass | 10,307 | 3,375 | -67% | 1 | 1 | 0% | 1,693 | 895 | -47% | 0 | 0 | — |
case-14 | fail→pass | 6,752 | 3,204 | -53% | 1 | 1 | 0% | 1,069 | 867 | -19% | 0 | 0 | — |
case-15 | pass→pass | 7,297 | 2,160 | -70% | 1 | 1 | 0% | 1,191 | 626 | -47% | 0 | 0 | — |
case-16 | fail→pass | 19,632 | 8,662 | -56% | 1 | 1 | 0% | 3,717 | 1,882 | -49% | 0 | 0 | — |
case-17 | pass→pass | 9,024 | 2,350 | -74% | 1 | 1 | 0% | 1,195 | 669 | -44% | 0 | 0 | — |
case-18 | fail→pass | 15,953 | 2,756 | -83% | 1 | 1 | 0% | 2,424 | 762 | -69% | 0 | 0 | — |
case-19 | fail→pass | 13,903 | 7,781 | -44% | 1 | 1 | 0% | 2,235 | 1,744 | -22% | 0 | 0 | — |
case-20 | pass→pass | 5,950 | 6,308 | +6% | 1 | 1 | 0% | 940 | 1,302 | +39% | 0 | 0 | — |
case-22 | pass→pass | 10,486 | 2,457 | -77% | 1 | 1 | 0% | 1,726 | 655 | -62% | 0 | 0 | — |
case-23 | fail→pass | 11,386 | 2,675 | -77% | 1 | 1 | 0% | 1,999 | 779 | -61% | 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 21 counted toward the lift figure. The other 2 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 +61 percentage points is the difference between those two pass rates over the 21 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.