Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Activate when you are a delegated subagent (not the orchestrator). Establishes subagent protocol with terse returns, details to history/, file ownership boundaries, and escalation rules. You implement; orchestrator reviews and commits.
.claude/skills/majiayu000-subagent/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-10 | ✗→✓ | ▲ Improved | -41% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 175% | 0% |
| case-12 | ✗→✓ | ▲ Improved | 43% | 0% |
| case-14 | ✗→✓ | ▲ Improved | 68% | 0% |
| case-15 | ✗→✓ | ▲ Improved | -24% | 0% |
You are a subagent — delegated by an orchestrator to implement a specific task. Your job is focused execution with minimal token footprint on return.
history/| Content | Destination | |---------|-------------| | Summary (1-5 lines) | Return to orchestrator | | Implementation details | history/<bead-id>.md or history/session.md | | Logs, traces, verbose output | history/ or /tmp/claude-* | | Capability gaps | Summary + history/gaps.log |
DONE: <what you completed>
CHANGED: <files modified>
RESULT: <pass/fail, test results if applicable>
BLOCKERS: <none, or what stopped you>
GAPS: <capabilities you wished you had>If history/ doesn't exist:
mkdir -p history && echo 'history/' >> .gitignore/tmp/claude-<project>-<date>.logYour prompt should include OWN and READ-ONLY lists.
| List | Permission | |------|------------| | OWN | Create, edit, delete freely | | READ-ONLY | Read only — do not modify | | Unlisted | Ask orchestrator before touching |
Never modify:
bd close, no status changes)Run verification commands specified in your prompt before returning.
Common gates:
npm run check (lint + typecheck + test)cargo check && cargo testgo build ./... && go test ./...If gates fail, fix and retry. If you can't fix, report in BLOCKERS.
Escalate immediately if:
How to escalate: Return summary with BLOCKERS section explaining what you need.
Activate skills specified in your prompt. Common ones:
typescript-pro, go-pro, rust-pro — language expertisesolid-architecture — design principlesgame-perf — hot-path optimization| Don't | Why | |-------|-----| | Return full file contents | Wastes orchestrator tokens | | Dump verbose logs | Put in history/ instead | | Commit changes | Orchestrator owns git | | Close beads | Orchestrator verifies first | | Modify unlisted files | Violates ownership boundaries | | Spin on blockers | Escalate after 3 attempts | | Hide failures | Report honestly in summary |
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-02 | fail→fail | 6,047 | 17,484 | +189% | 1 | 1 | 0% | 593 | 1,297 | +119% | 0 | 0 | — |
case-01 | fail→fail | 10,428 | 10,836 | +4% | 1 | 1 | 0% | 339 | 1,138 | +236% | 0 | 0 | — |
case-03 | fail→fail | 13,202 | 10,424 | -21% | 1 | 1 | 0% | 324 | 1,218 | +276% | 0 | 0 | — |
case-04 | pass→fail | 33,559 | 4,511 | -87% | 1 | 1 | 0% | 4,514 | 1,013 | -78% | 0 | 0 | — |
case-05 | fail→fail | 26,841 | 14,907 | -44% | 1 | 1 | 0% | 214 | 3,395 | +1486% | 0 | 0 | — |
case-06 | pass→fail | 32,180 | 27,870 | -13% | 1 | 1 | 0% | 4,160 | 2,347 | -44% | 0 | 0 | — |
case-07 | pass→pass | 11,265 | 17,816 | +58% | 1 | 1 | 0% | 1,001 | 1,982 | +98% | 0 | 0 | — |
case-08 | fail→fail | 13,059 | 13,347 | +2% | 1 | 1 | 0% | 1,202 | 1,318 | +10% | 0 | 0 | — |
case-09 | fail→fail | 26,256 | 11,760 | -55% | 1 | 1 | 0% | 1,881 | 1,217 | -35% | 0 | 0 | — |
case-10 | fail→pass | 11,074 | 2,660 | -76% | 1 | 1 | 0% | 1,889 | 1,122 | -41% | 0 | 0 | — |
case-11 | fail→pass | 7,271 | 7,521 | +3% | 1 | 1 | 0% | 420 | 1,153 | +175% | 0 | 0 | — |
case-12 | fail→pass | 17,287 | 8,410 | -51% | 1 | 1 | 0% | 928 | 1,330 | +43% | 0 | 0 | — |
case-13 | pass→pass | 15,455 | 8,875 | -43% | 1 | 1 | 0% | 1,089 | 1,412 | +30% | 0 | 0 | — |
case-14 | fail→pass | 10,337 | 7,574 | -27% | 1 | 1 | 0% | 765 | 1,283 | +68% | 0 | 0 | — |
case-15 | fail→pass | 16,095 | 4,921 | -69% | 1 | 1 | 0% | 1,847 | 1,397 | -24% | 0 | 0 | — |
case-16 | fail→fail | 18,982 | 7,497 | -61% | 1 | 1 | 0% | 1,606 | 1,125 | -30% | 0 | 0 | — |
case-17 | fail→pass | 16,815 | 3,494 | -79% | 1 | 1 | 0% | 2,011 | 1,317 | -35% | 0 | 0 | — |
case-18 | fail→fail | 11,102 | 2,867 | -74% | 1 | 1 | 0% | 1,104 | 1,171 | +6% | 0 | 0 | — |
case-23 | fail→pass | 10,198 | 7,249 | -29% | 1 | 1 | 0% | 1,797 | 1,079 | -40% | 0 | 0 | — |
case-19 | fail→fail | 10,931 | 2,999 | -73% | 1 | 1 | 0% | 971 | 1,299 | +34% | 0 | 0 | — |
case-20 | pass→pass | 16,481 | 9,484 | -42% | 1 | 1 | 0% | 1,771 | 1,412 | -20% | 0 | 0 | — |
case-21 | pass→pass | 3,268 | 9,143 | +180% | 1 | 1 | 0% | 455 | 1,464 | +222% | 0 | 0 | — |
case-22 | fail→pass | 10,636 | 1,974 | -81% | 1 | 1 | 0% | 1,486 | 1,045 | -30% | 0 | 0 | — |
case-24 | fail→pass | 11,330 | 2,154 | -81% | 1 | 1 | 0% | 961 | 1,092 | +14% | 0 | 0 | — |
case-25 | pass→pass | 15,668 | 11,344 | -28% | 1 | 1 | 0% | 2,319 | 1,802 | -22% | 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 18 counted toward the lift figure. The other 7 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 +28 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.
Other measured skills in the registry, with their headline benchmark lift.