Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →GitHub releases command center -- create, edit, and manage releases and their binary assets entirely from the editor. Bypasses the drag-and-drop asset upload and icon-only controls that are inaccessible to screen readers.
.claude/skills/community-access-release-manager/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | 17% | 0% |
| case-01 | ✓→✗ | ▼ Worse | -17% | 0% |
| case-11 | ✓→✗ | ▼ Worse | 7% | 0% |
| case-09 | ✓→✗ | ▼ Worse | 101% | 0% |
| case-22 | ✓→✗ | ▼ Worse | 7% | 0% |
Derived from .claude/agents/release-manager.md. Treat platform-specific tool names or delegation instructions as Codex equivalents.
Shared instructions
Skills: github-workflow-standards, github-scanning
You are the Release Manager. You give screen reader users and keyboard-only users full control over GitHub releases and binary assets — a feature whose web UI relies on drag-and-drop file upload zones, icon-only delete buttons with inconsistent labels, and the Monaco markdown editor.
GitHub's release management UI presents accessibility barriers:
gh api user.gh release CLI. Never instruct the user to use the web upload UI.| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | pass→fail | 8,522 | 7,156 | -16% | 1 | 1 | 0% | 1,513 | 1,258 | -17% | 0 | 0 | — |
case-02 | fail→fail | 3,430 | 7,512 | +119% | 1 | 1 | 0% | 402 | 1,187 | +195% | 0 | 0 | — |
case-03 | fail→fail | 9,145 | 4,290 | -53% | 1 | 1 | 0% | 1,715 | 892 | -48% | 0 | 0 | — |
case-04 | pass→pass | 3,430 | 4,843 | +41% | 1 | 1 | 0% | 546 | 1,469 | +169% | 0 | 0 | — |
case-05 | pass→pass | 5,434 | 2,884 | -47% | 1 | 1 | 0% | 936 | 1,233 | +32% | 0 | 0 | — |
case-06 | pass→pass | 4,031 | 3,463 | -14% | 1 | 1 | 0% | 605 | 1,229 | +103% | 0 | 0 | — |
case-11 | pass→fail | 5,297 | 5,818 | +10% | 1 | 1 | 0% | 890 | 954 | +7% | 0 | 0 | — |
case-07 | fail→pass | 7,858 | 5,773 | -27% | 1 | 1 | 0% | 1,538 | 1,797 | +17% | 0 | 0 | — |
case-08 | pass→pass | 4,046 | 3,988 | -1% | 1 | 1 | 0% | 743 | 1,427 | +92% | 0 | 0 | — |
case-09 | pass→fail | 3,350 | 4,455 | +33% | 1 | 1 | 0% | 525 | 1,053 | +101% | 0 | 0 | — |
case-10 | fail→fail | 17,837 | 7,135 | -60% | 1 | 1 | 0% | 3,746 | 1,073 | -71% | 0 | 0 | — |
case-12 | pass→pass | 4,204 | 1,616 | -62% | 1 | 1 | 0% | 598 | 921 | +54% | 0 | 0 | — |
case-13 | pass→pass | 7,202 | 7,742 | +7% | 1 | 1 | 0% | 1,334 | 2,147 | +61% | 0 | 0 | — |
case-14 | pass→pass | 10,617 | 7,967 | -25% | 1 | 1 | 0% | 2,092 | 2,273 | +9% | 0 | 0 | — |
case-15 | pass→pass | 4,870 | 3,555 | -27% | 1 | 1 | 0% | 811 | 1,209 | +49% | 0 | 0 | — |
case-16 | pass→pass | 4,718 | 5,856 | +24% | 1 | 1 | 0% | 812 | 1,723 | +112% | 0 | 0 | — |
case-17 | fail→fail | 9,439 | 3,981 | -58% | 1 | 1 | 0% | 1,757 | 972 | -45% | 0 | 0 | — |
case-18 | fail→fail | 13,489 | 4,697 | -65% | 1 | 1 | 0% | 2,660 | 1,330 | -50% | 0 | 0 | — |
case-19 | pass→pass | 10,263 | 7,562 | -26% | 1 | 1 | 0% | 1,937 | 1,990 | +3% | 0 | 0 | — |
case-20 | pass→pass | 4,936 | 4,663 | -6% | 1 | 1 | 0% | 796 | 1,508 | +89% | 0 | 0 | — |
case-21 | fail→fail | 10,344 | 11,418 | +10% | 1 | 1 | 0% | 527 | 961 | +82% | 0 | 0 | — |
case-22 | pass→fail | 5,304 | 5,729 | +8% | 1 | 1 | 0% | 996 | 1,066 | +7% | 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 13 counted toward the lift figure. The other 9 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 -14 percentage points is the difference between those two pass rates over the 13 comparable cases. 6 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.