Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Orchestrate a comprehensive git workflow from code review through PR creation, leveraging specialized agents for quality assurance, testing, and deployment readiness. This workflow implements modern g
.claude/skills/git-pr-workflows-git-workflow/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-08 | ✗→✓ | ▲ Improved | 39% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 75% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 58% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 11% | 0% |
| case-20 | ✗→✓ | ▲ Improved | 15% | 0% |
Move completed changes from local review to a verified pull request without bypassing repository policy or branch protection.
Use for completed implementation work that must be reviewed, tested, committed, pushed to a topic branch, and opened as a pull request. Use the repository's dedicated maintainer or release workflow instead when one is mandatory.
Before mutation:
AGENTS.md, contribution guidance, maintainer docs, and relevant nested instructions.Repository policy wins over flags and user shorthand. Trunk-based development does not imply a direct push: when the target is protected, use a short-lived branch and pull request. If a repository defines a mandatory maintainer skill or guarded merge command, hand off merge and release actions to it. In agentic-awesome-skills, use antigravity-maintainer-batch-release and npm run merge:batch.
Resolve these from the request and repository:
Ask only when a missing choice changes the result materially.
bashgit status --short --branch git diff --stat git diff --cached --stat git branch --show-current git remote -v
Confirm every file in scope. Stop if staged or dirty files cannot be separated safely.
When subagents are available and authorized, assign independent bounded passes for:
Give each reviewer the raw diff and repository instructions. Keep the main agent responsible for deduplication, severity, edits, and final verification.
Run the repository's targeted checks, then its required pre-PR suite. If a check fails:
Do not weaken gates, hide skipped tests, or treat deterministic failures as flaky.
Fetch the target before committing. If currently on a protected/default branch, create a topic branch before mutation.
bashgit fetch origin <target-branch> git switch -c <topic-branch> origin/<target-branch> git status --short --branch
Stage only intended paths and create focused conventional commits according to repository policy. Rebase or update the topic branch when strict required checks demand the latest target; never force a shared branch without explicit authorization.
bashgit push -u origin <topic-branch> gh pr create --base <target-branch> --head <topic-branch> \ --title "<conventional title>" --body-file <body-file>
The PR body must truthfully include:
Never mark a pending automated review or test as completed.
bashgh pr view <pr-number> --json headRefOid,baseRefOid,mergeable,mergeStateStatus,url gh pr checks <pr-number>
Bind review evidence to the current full head SHA. If the head or base changes, discard stale conclusions and rerun affected checks.
Use the repository's guarded merge path. Do not replace required checks, merge queues, exact-SHA attestations, or maintainer commands with a raw merge API. After merge, fetch the target and verify the requested remote, CI, deployment, or release state.
Finish when the PR exists at the intended head, required checks are green or have one exact blocker, review evidence is current, unrelated user work is preserved, and any requested guarded merge or deployment is verified.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-08 | fail→pass | 9,104 | 7,397 | -19% | 1 | 1 | 0% | 1,653 | 2,298 | +39% | 0 | 0 | — |
case-01 | fail→fail | 4,291 | 7,270 | +69% | 1 | 1 | 0% | 249 | 1,585 | +537% | 0 | 0 | — |
case-02 | fail→fail | 15,132 | 4,557 | -70% | 1 | 1 | 0% | 3,171 | 1,263 | -60% | 0 | 0 | — |
case-03 | fail→fail | 3,013 | 4,936 | +64% | 1 | 1 | 0% | 322 | 1,309 | +307% | 0 | 0 | — |
case-04 | fail→pass | 6,470 | 7,296 | +13% | 1 | 1 | 0% | 1,315 | 2,305 | +75% | 0 | 0 | — |
case-05 | fail→fail | 22,645 | 14,935 | -34% | 1 | 1 | 0% | 818 | 1,382 | +69% | 0 | 0 | — |
case-06 | pass→pass | 8,025 | 10,224 | +27% | 1 | 1 | 0% | 1,304 | 2,031 | +56% | 0 | 0 | — |
case-07 | pass→pass | 15,832 | 7,919 | -50% | 1 | 1 | 0% | 2,541 | 2,398 | -6% | 0 | 0 | — |
case-09 | pass→pass | 9,890 | 3,758 | -62% | 1 | 1 | 0% | 1,435 | 1,613 | +12% | 0 | 0 | — |
case-10 | pass→pass | 10,577 | 3,102 | -71% | 1 | 1 | 0% | 1,687 | 1,684 | -0% | 0 | 0 | — |
case-11 | fail→fail | 6,795 | 4,714 | -31% | 1 | 1 | 0% | 1,232 | 1,880 | +53% | 0 | 0 | — |
case-12 | pass→pass | 8,164 | 6,842 | -16% | 1 | 1 | 0% | 1,529 | 2,267 | +48% | 0 | 0 | — |
case-13 | fail→pass | 5,770 | 3,617 | -37% | 1 | 1 | 0% | 1,109 | 1,751 | +58% | 0 | 0 | — |
case-14 | pass→pass | 8,713 | 5,757 | -34% | 1 | 1 | 0% | 1,510 | 1,981 | +31% | 0 | 0 | — |
case-15 | fail→pass | 10,488 | 4,863 | -54% | 1 | 1 | 0% | 1,617 | 1,800 | +11% | 0 | 0 | — |
case-16 | pass→pass | 5,377 | 2,968 | -45% | 1 | 1 | 0% | 904 | 1,491 | +65% | 0 | 0 | — |
case-17 | fail→fail | 8,944 | 2,807 | -69% | 1 | 1 | 0% | 1,696 | 1,568 | -8% | 0 | 0 | — |
case-18 | pass→pass | 10,742 | 5,858 | -45% | 1 | 1 | 0% | 1,558 | 1,967 | +26% | 0 | 0 | — |
case-19 | pass→pass | 8,511 | 4,547 | -47% | 1 | 1 | 0% | 1,253 | 1,834 | +46% | 0 | 0 | — |
case-20 | fail→pass | 10,660 | 4,357 | -59% | 1 | 1 | 0% | 1,617 | 1,854 | +15% | 0 | 0 | — |
case-21 | pass→pass | 10,278 | 7,058 | -31% | 1 | 1 | 0% | 1,560 | 2,394 | +53% | 0 | 0 | — |
case-22 | pass→pass | 12,849 | 2,162 | -83% | 1 | 1 | 0% | 2,100 | 1,364 | -35% | 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 18 counted toward the lift figure. The other 4 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 +23 percentage points is the difference between those two pass rates over the 18 comparable cases. 2 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.
| Model | Method | Date | Lift |
|---|---|---|---|
| gemini-3.6-flash | verified | 7/28/2026 | +41% |
Other measured skills in the registry, with their headline benchmark lift.