Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Reviews Home Assistant GitHub pull requests and provides feedback comments. This is the top skill to use for reviewing Pull Requests from GitHub.
.claude/skills/home-assistant-ha-pr-reviewer/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | 5% | 0% |
| case-09 | ✓→✗ | ▼ Worse | -89% | 0% |
| case-23 | ✓→✗ | ▼ Worse | -59% | 0% |
| case-25 | ✓→✗ | ▼ Worse | -69% | 0% |
| case-01 | ✗→✗ | ＝ Same ✗ | -93% | 0% |
ha-review skill. It is VERY IMPORTANT to follow the ha-review skill instructions. Explicitly pass the PR's target/base branch to the ha-review skill (obtained via gh pr view) so it diffs against the correct base.ha-pr-comment-audit skill.| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 73,762 | 14,980 | -80% | 1 | 1 | 0% | 5,631 | 387 | -93% | 0 | 0 | — |
case-02 | fail→fail | 25,165 | 17,078 | -32% | 1 | 1 | 0% | 3,181 | 459 | -86% | 0 | 0 | — |
case-03 | fail→fail | 17,532 | 11,071 | -37% | 1 | 1 | 0% | 2,593 | 441 | -83% | 0 | 0 | — |
case-04 | fail→fail | 13,142 | 126,491 | +862% | 1 | 1 | 0% | 1,936 | 410 | -79% | 0 | 0 | — |
case-05 | fail→fail | 5,385 | 25,564 | +375% | 1 | 1 | 0% | 767 | 594 | -23% | 0 | 0 | — |
case-06 | fail→fail | 45,669 | 6,621 | -86% | 1 | 1 | 0% | 479 | 496 | +4% | 0 | 0 | — |
case-07 | fail→pass | 12,587 | 19,257 | +53% | 1 | 1 | 0% | 1,882 | 1,970 | +5% | 0 | 0 | — |
case-08 | fail→fail | 22,199 | 9,078 | -59% | 1 | 1 | 0% | 320 | 564 | +76% | 0 | 0 | — |
case-09 | pass→fail | 40,325 | 8,326 | -79% | 1 | 1 | 0% | 4,224 | 476 | -89% | 0 | 0 | — |
case-10 | fail→fail | 17,540 | 4,879 | -72% | 1 | 1 | 0% | 2,422 | 430 | -82% | 0 | 0 | — |
case-11 | fail→fail | 19,925 | 5,568 | -72% | 1 | 1 | 0% | 2,832 | 404 | -86% | 0 | 0 | — |
case-12 | fail→fail | 16,805 | 7,995 | -52% | 1 | 1 | 0% | 849 | 511 | -40% | 0 | 0 | — |
case-13 | fail→fail | 14,975 | 6,855 | -54% | 1 | 1 | 0% | 2,045 | 520 | -75% | 0 | 0 | — |
case-14 | fail→fail | 53,370 | 6,940 | -87% | 1 | 1 | 0% | 4,750 | 422 | -91% | 0 | 0 | — |
case-15 | fail→fail | 4,237 | 7,524 | +78% | 1 | 1 | 0% | 587 | 554 | -6% | 0 | 0 | — |
case-16 | fail→fail | 9,639 | 5,554 | -42% | 1 | 1 | 0% | 1,495 | 402 | -73% | 0 | 0 | — |
case-17 | fail→fail | 12,954 | 9,124 | -30% | 1 | 1 | 0% | 1,881 | 763 | -59% | 0 | 0 | — |
case-18 | fail→fail | 17,756 | 10,601 | -40% | 1 | 1 | 0% | 2,249 | 474 | -79% | 0 | 0 | — |
case-19 | fail→fail | 19,147 | 5,526 | -71% | 1 | 1 | 0% | 2,565 | 374 | -85% | 0 | 0 | — |
case-20 | fail→fail | 14,194 | 7,612 | -46% | 1 | 1 | 0% | 2,334 | 500 | -79% | 0 | 0 | — |
case-21 | fail→fail | 11,100 | 8,271 | -25% | 1 | 1 | 0% | 1,598 | 562 | -65% | 0 | 0 | — |
case-22 | fail→fail | 4,763 | 7,799 | +64% | 1 | 1 | 0% | 580 | 493 | -15% | 0 | 0 | — |
case-23 | pass→fail | 9,121 | 11,348 | +24% | 1 | 1 | 0% | 1,398 | 574 | -59% | 0 | 0 | — |
case-24 | fail→fail | 13,440 | 7,998 | -40% | 1 | 1 | 0% | 614 | 448 | -27% | 0 | 0 | — |
case-25 | pass→fail | 10,697 | 7,882 | -26% | 1 | 1 | 0% | 1,579 | 483 | -69% | 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. 25 cases were attempted, and 1 counted toward the lift figure. The other 24 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 -8 percentage points is the difference between those two pass rates over the 1 comparable cases. 10 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.