Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Execute the plan by dispatching fresh subagents per task, monitoring status, and collecting results
.claude/skills/yogsoth-ai-experiment-running/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | -6% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 1330% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -7% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -28% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 6% | 0% |
Key Question: How to execute?
实现的执行直接交给 superpowers 现成链路，不再自写 fresh-subagent / 三段 review。 plan（上游 plan-writing 产出）就绪后，本策略是一串决策节点：
Skill load superpowers:using-git-worktrees —— 建隔离工作区 + 跑 baseline 测试。Skill load ponytail:ponytail —— 进入写代码前开启精简反射（边写边 lean）。多数实验实现用前者；任务高度独立、需强隔离时用后者（详见 subagent-execution-loop）。
Skill load superpowers:verification-before-completion —— claim 完成前先跑证明命令。Skill load ponytail:ponytail-debt —— 收尾前收集 ponytail: 欠债标记。Skill load superpowers:finishing-a-development-branch —— 验证测试 → merge/PR/branch。DARE 原生的 checkpoint-and-recover（高风险操作前存档）与 subagent-execution-loop （执行循环细节）作为 tactic 仍在编排内保留。
[plan from plan-writing]
→ superpowers:using-git-worktrees (隔离区 + baseline)
→ ponytail:ponytail (精简反射开启)
→ superpowers:executing-plans 或 superpowers:subagent-driven-development
→ superpowers:verification-before-completion (claim 前验证)
→ ponytail:ponytail-debt (收欠债)
→ superpowers:finishing-a-development-branch (收尾)| Step | Max Budget | Output | |------|-----------|--------| | Per-task execution | 50% of execution budget / N tasks | Task result | | Monitoring overhead | 5% of execution budget | Status log | | Retry budget | 10% of execution budget | Unblocked tasks |
<!-- BEGIN available-tables (generated) --> <!-- external rows hand-maintained; do not regenerate this file -->
Optional, no fixed order; the final leaf is always a sop.
| Tactic | When to use | | --- | --- | | checkpoint-and-recover | Checkpoint state before risky operations, detect anomalies, and recover gracefully | | subagent-execution-loop | Orchestrate task execution via fresh subagents with dispatch, monitoring, and result collection |
Optional, no fixed order; the final leaf is always a sop.
| SOP | When to use | | --- | --- | | execution-monitoring | Monitor execution progress, detect anomalies, and report status | | implementer-dispatch | Dispatch execution subagent — select model by complexity, construct prompt with full task context | | ponytail:ponytail | Lazy-senior reflex: simplest thing that holds; mark every deliberate shortcut | | ponytail:ponytail-debt | Harvest ponytail debt markers before finishing | | result-collection | Collect experiment outputs — metrics, logs, artifacts — into structured result set | | superpowers:executing-plans | Execute the plan task-by-task in the current session with checkpoints | | superpowers:finishing-a-development-branch | Verify tests -> merge / PR / branch cleanup | | superpowers:subagent-driven-development | Execute the plan via a fresh subagent per task with two-stage review | | superpowers:using-git-worktrees | Create an isolated worktree + run baseline tests before implementing | | superpowers:verification-before-completion | Run the proving command and confirm output before claiming done |
<!-- END available-tables (generated) -->
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 2,805 | 11,735 | +318% | 1 | 1 | 0% | 125 | 1,368 | +994% | 0 | 0 | — |
case-02 | fail→pass | 14,449 | 5,957 | -59% | 1 | 1 | 0% | 2,261 | 2,115 | -6% | 0 | 0 | — |
case-03 | fail→pass | 29,567 | 6,704 | -77% | 1 | 1 | 0% | 160 | 2,288 | +1330% | 0 | 0 | — |
case-04 | fail→pass | 9,250 | 2,375 | -74% | 1 | 1 | 0% | 1,531 | 1,421 | -7% | 0 | 0 | — |
case-05 | fail→pass | 14,042 | 3,389 | -76% | 1 | 1 | 0% | 2,302 | 1,664 | -28% | 0 | 0 | — |
case-06 | fail→pass | 7,486 | 1,775 | -76% | 1 | 1 | 0% | 1,201 | 1,278 | +6% | 0 | 0 | — |
case-07 | fail→pass | 7,017 | 2,413 | -66% | 1 | 1 | 0% | 1,021 | 1,405 | +38% | 0 | 0 | — |
case-08 | fail→pass | 5,277 | 2,895 | -45% | 1 | 1 | 0% | 787 | 1,483 | +88% | 0 | 0 | — |
case-09 | fail→pass | 13,091 | 2,512 | -81% | 1 | 1 | 0% | 1,017 | 1,445 | +42% | 0 | 0 | — |
case-10 | pass→pass | 16,413 | 1,982 | -88% | 1 | 1 | 0% | 1,265 | 1,298 | +3% | 0 | 0 | — |
case-11 | pass→pass | 24,378 | 1,908 | -92% | 1 | 1 | 0% | 576 | 1,255 | +118% | 0 | 0 | — |
case-12 | pass→pass | 6,751 | 2,444 | -64% | 1 | 1 | 0% | 1,051 | 1,347 | +28% | 0 | 0 | — |
case-13 | pass→pass | 7,978 | 1,822 | -77% | 1 | 1 | 0% | 1,227 | 1,256 | +2% | 0 | 0 | — |
case-14 | fail→pass | 7,100 | 1,662 | -77% | 1 | 1 | 0% | 1,120 | 1,243 | +11% | 0 | 0 | — |
case-15 | pass→fail | 6,752 | 2,003 | -70% | 1 | 1 | 0% | 1,016 | 1,362 | +34% | 0 | 0 | — |
case-16 | fail→pass | 7,376 | 2,590 | -65% | 1 | 1 | 0% | 1,184 | 1,524 | +29% | 0 | 0 | — |
case-17 | fail→pass | 10,601 | 2,208 | -79% | 1 | 1 | 0% | 1,549 | 1,353 | -13% | 0 | 0 | — |
case-18 | pass→pass | 7,962 | 2,405 | -70% | 1 | 1 | 0% | 1,160 | 1,398 | +21% | 0 | 0 | — |
case-19 | fail→pass | 12,520 | 2,224 | -82% | 1 | 1 | 0% | 1,963 | 1,323 | -33% | 0 | 0 | — |
case-20 | pass→pass | 30,974 | 32,181 | +4% | 1 | 1 | 0% | 5,656 | 7,156 | +27% | 0 | 0 | — |
case-21 | fail→fail | 6,687 | 2,412 | -64% | 1 | 1 | 0% | 1,172 | 1,282 | +9% | 0 | 0 | — |
case-22 | pass→pass | 17,712 | 20,354 | +15% | 1 | 1 | 0% | 2,802 | 4,127 | +47% | 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 20 counted toward the lift figure. The other 2 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 +50 percentage points is the difference between those two pass rates over the 20 comparable cases. 1 case got worse with the skill loaded, and it is 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.