Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Run controlled, paired evaluations of an Agent Skill against a no-skill baseline or previous revision, using isolated contexts, raw artifacts, deterministic verification, and blind comparison. Use when validating a new or edited skill, testing whether instructions improve pass rate enough to justify tokens or latency, resolving competing skill versions, or producing reproducible efficacy evidence. Do not use for a quick structural lint or an upgrade-only smoke check.
.claude/skills/paranoidandroid2124-trust-me-bro/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | -28% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 10% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 137% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -36% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -24% | 0% |
Measure the marginal value of one skill. Keep the task and environment fixed; change only the skill condition.
“Trust me” is not a control condition.
reasoning setting, verifier, and timeout.
tasks for a smoke trial. Require repeated runs before publishing a performance claim.
baseline: no candidate skill content;candidate: the selected skill revision;previous: optional prior revision.record and blind comparator contract.
workers discover the candidate skill or prior conclusions.
budget. Give only candidate workers the candidate skill.
user interventions. Do not grade from worker self-reports.
exact target. Prefer fixtures, sandboxes, or drafts.
them to a fresh comparator. The comparator receives the original task and rubric, but no condition labels, expected winner, or implementer analysis.
grading it.
bash node scripts/summarize-trials.mjs results.jsonl node scripts/summarize-trials.mjs results.jsonl --reference previous --comparison candidate
and intervention cost.
regression.
One run can catch a bug; it cannot establish a stable uplift. State sample size and uncertainty without manufacturing statistical confidence.
Return the pinned configuration, task matrix, condition-level results, paired improvements/regressions, cost delta, raw artifact locations, decision, and known limits. Never market a smoke test as a benchmark.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 21,810 | 34,904 | +60% | 1 | 1 | 0% | 4,098 | 6,828 | +67% | 0 | 0 | — |
case-02 | fail→pass | 31,981 | 19,683 | -38% | 1 | 1 | 0% | 5,700 | 4,096 | -28% | 0 | 0 | — |
case-03 | fail→pass | 22,576 | 21,209 | -6% | 1 | 1 | 0% | 4,252 | 4,687 | +10% | 0 | 0 | — |
case-04 | pass→pass | 13,571 | 5,637 | -58% | 1 | 1 | 0% | 2,238 | 1,425 | -36% | 0 | 0 | — |
case-05 | fail→pass | 31,954 | 14,438 | -55% | 1 | 1 | 0% | 1,217 | 2,886 | +137% | 0 | 0 | — |
case-06 | fail→fail | 11,106 | 6,418 | -42% | 1 | 1 | 0% | 1,963 | 1,615 | -18% | 0 | 0 | — |
case-07 | pass→pass | 14,300 | 7,831 | -45% | 1 | 1 | 0% | 2,421 | 1,873 | -23% | 0 | 0 | — |
case-08 | fail→pass | 12,322 | 4,694 | -62% | 1 | 1 | 0% | 2,056 | 1,314 | -36% | 0 | 0 | — |
case-09 | pass→pass | 13,399 | 3,253 | -76% | 1 | 1 | 0% | 1,847 | 1,063 | -42% | 0 | 0 | — |
case-10 | pass→pass | 14,137 | 5,871 | -58% | 1 | 1 | 0% | 2,223 | 1,499 | -33% | 0 | 0 | — |
case-11 | fail→pass | 13,887 | 5,284 | -62% | 1 | 1 | 0% | 1,921 | 1,451 | -24% | 0 | 0 | — |
case-12 | pass→pass | 9,933 | 5,153 | -48% | 1 | 1 | 0% | 1,599 | 1,390 | -13% | 0 | 0 | — |
case-13 | pass→pass | 13,850 | 5,342 | -61% | 1 | 1 | 0% | 2,394 | 1,438 | -40% | 0 | 0 | — |
case-14 | pass→pass | 10,459 | 4,440 | -58% | 1 | 1 | 0% | 1,781 | 1,276 | -28% | 0 | 0 | — |
case-15 | fail→pass | 11,737 | 5,324 | -55% | 1 | 1 | 0% | 1,940 | 1,399 | -28% | 0 | 0 | — |
case-16 | fail→pass | 21,350 | 1,877 | -91% | 1 | 1 | 0% | 1,799 | 924 | -49% | 0 | 0 | — |
case-17 | pass→pass | 11,808 | 7,971 | -32% | 1 | 1 | 0% | 1,856 | 1,819 | -2% | 0 | 0 | — |
case-18 | pass→pass | 9,606 | 5,257 | -45% | 1 | 1 | 0% | 1,626 | 1,467 | -10% | 0 | 0 | — |
case-19 | pass→pass | 6,648 | 3,885 | -42% | 1 | 1 | 0% | 1,000 | 1,066 | +7% | 0 | 0 | — |
case-20 | pass→pass | 10,552 | 5,208 | -51% | 1 | 1 | 0% | 1,755 | 1,440 | -18% | 0 | 0 | — |
case-21 | pass→pass | 16,100 | 16,655 | +3% | 1 | 1 | 0% | 2,738 | 3,062 | +12% | 0 | 0 | — |
case-22 | pass→pass | 4,845 | 4,247 | -12% | 1 | 1 | 0% | 1,041 | 1,384 | +33% | 0 | 0 | — |
case-23 | pass→pass | 11,377 | 7,917 | -30% | 1 | 1 | 0% | 2,312 | 2,198 | -5% | 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. 23 cases were attempted, and 22 counted toward the lift figure. The other 1 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 +30 percentage points is the difference between those two pass rates over the 22 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.