Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when an FRB PR contains independently landable work that should be extracted into predecessor PRs and connected as a GitHub PR chain without wasting full automatic CI.
.claude/skills/fzyzcjy-frb-pr-chain-split/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | 26% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 20% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 48% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 42% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 10% | 0% |
flutter_rust_bridge PR when part of its diff is independent of the PR's stated objective.Classify every suspicious change before moving commits:
| Classification | Destination | Reason | | --- | --- | --- | | Independent bug fix or hardening | Predecessor PR | It can land and provide value before the main task | | Independent mechanical refactor | Predecessor PR | It removes review noise from the main task | | Main-task prerequisite with its own contract | Predecessor PR | The main PR can depend on a separately reviewable foundation | | Generated output caused by the main task | Main PR | It cannot land meaningfully by itself | | Test or documentation for the main behavior | Main PR | Splitting would separate the contract from its implementation |
frb-prepare-pr for generation, lint, test, and review requirements.textmaster <- predecessor-a <- predecessor-b <- main-task
github/gh-stack extension through gh stack.gh stack for every stack operation: initialization, restructuring, linking, submission, synchronization, rebasing, pushing, and verification.gh pr create --base, gh pr edit --base, direct API base edits, or branch ancestry alone.gh stack creates or updates the native GitHub stack object.gh 2.0 or later; a missing gh stack command normally means the extension is absent, not that core gh is outdated.bashgh stack --help gh extension install github/gh-stack
gh stack init and gh stack submit.--force-with-lease only after those checks when published history must be rewritten.gh stack link. It pushes branches, creates missing PRs, corrects their base branches, and creates the native GitHub stack object.bashgh stack init --base master <predecessor-branch> <main-task-branch> gh stack submit
bashgh stack link --base master <predecessor-branch> <main-task-branch>
bashgit merge-base --is-ancestor <predecessor-branch> <main-task-branch> gh stack view --json
<previous-branch>...<predecessor-branch> contains only that predecessor;<final-predecessor-branch>...<main-task-branch> no longer contains the extracted work;frb-ci-filter before manipulating CI.ci-manual-dispatch after creating every predecessor PR:bashgh pr edit <predecessor-pr-number> --add-label ci-manual-dispatch
pull_request:labeled run. The workflow concurrency group cancels the earlier automatic PR run, and the labeled run produces an empty heavy-job plan.ci-manual-dispatch on the PR;bash./frb_internal plan-ci --filter '<filter>' gh workflow run ci.yaml --ref <predecessor-branch> -f 'ci_filter=<filter>'
ci-manual-dispatch and obtain the normal required CI unless the maintainer explicitly accepts narrower evidence.ci-manual-dispatch while it is not awaiting final full CI.ci_filter.gh stack link or gh stack submit.gh stack view --json.ci-manual-dispatch on the PR when claiming normal full-CI readiness.| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 25,344 | 18,342 | -28% | 1 | 1 | 0% | 4,592 | 4,831 | +5% | 0 | 0 | — |
case-02 | fail→pass | 23,742 | 20,900 | -12% | 1 | 1 | 0% | 4,208 | 5,284 | +26% | 0 | 0 | — |
case-03 | fail→pass | 21,171 | 15,148 | -28% | 1 | 1 | 0% | 3,554 | 4,262 | +20% | 0 | 0 | — |
case-04 | pass→pass | 17,009 | 10,133 | -40% | 1 | 1 | 0% | 2,476 | 3,236 | +31% | 0 | 0 | — |
case-05 | pass→pass | 13,510 | 7,955 | -41% | 1 | 1 | 0% | 2,025 | 3,057 | +51% | 0 | 0 | — |
case-06 | fail→pass | 10,398 | 4,237 | -59% | 1 | 1 | 0% | 1,532 | 2,268 | +48% | 0 | 0 | — |
case-07 | fail→pass | 11,241 | 6,028 | -46% | 1 | 1 | 0% | 1,783 | 2,535 | +42% | 0 | 0 | — |
case-08 | fail→fail | 13,603 | 8,138 | -40% | 1 | 1 | 0% | 2,036 | 2,988 | +47% | 0 | 0 | — |
case-09 | fail→pass | 15,152 | 6,170 | -59% | 1 | 1 | 0% | 2,430 | 2,674 | +10% | 0 | 0 | — |
case-10 | fail→pass | 12,652 | 7,739 | -39% | 1 | 1 | 0% | 1,813 | 2,742 | +51% | 0 | 0 | — |
case-11 | fail→pass | 11,222 | 5,739 | -49% | 1 | 1 | 0% | 2,071 | 2,581 | +25% | 0 | 0 | — |
case-12 | fail→pass | 11,418 | 4,608 | -60% | 1 | 1 | 0% | 1,985 | 2,347 | +18% | 0 | 0 | — |
case-13 | pass→pass | 14,283 | 7,420 | -48% | 1 | 1 | 0% | 2,108 | 2,883 | +37% | 0 | 0 | — |
case-14 | pass→pass | 11,291 | 7,049 | -38% | 1 | 1 | 0% | 1,535 | 2,924 | +90% | 0 | 0 | — |
case-15 | pass→pass | 14,757 | 11,767 | -20% | 1 | 1 | 0% | 2,700 | 3,333 | +23% | 0 | 0 | — |
case-16 | pass→pass | 12,790 | 3,846 | -70% | 1 | 1 | 0% | 1,051 | 2,158 | +105% | 0 | 0 | — |
case-17 | fail→pass | 10,632 | 5,592 | -47% | 1 | 1 | 0% | 1,632 | 2,369 | +45% | 0 | 0 | — |
case-18 | pass→pass | 6,382 | 4,108 | -36% | 1 | 1 | 0% | 982 | 2,136 | +118% | 0 | 0 | — |
case-19 | pass→pass | 14,257 | 8,776 | -38% | 1 | 1 | 0% | 2,271 | 2,795 | +23% | 0 | 0 | — |
case-20 | pass→pass | 14,462 | 15,011 | +4% | 1 | 1 | 0% | 2,173 | 3,933 | +81% | 0 | 0 | — |
case-21 | pass→pass | 15,592 | 20,126 | +29% | 1 | 1 | 0% | 2,930 | 5,023 | +71% | 0 | 0 | — |
case-22 | pass→pass | 13,879 | 15,934 | +15% | 1 | 1 | 0% | 2,087 | 4,283 | +105% | 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.
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.