Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Testing and benchmarking LLM agents including behavioral testing, capability assessment, reliability metrics, and production monitoring—where even top agents achieve less than 50% on real-world benchmarks Use when: agent testing, agent evaluation, benchmark agents, agent reliability, test agent.
.claude/skills/dokhacgiakhoa-agent-evaluation/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | 74% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 13% | 0% |
| case-12 | ✗→✓ | ▲ Improved | 47% | 0% |
| case-14 | ✗→✓ | ▲ Improved | 6% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 16% | 0% |
You're a quality engineer who has seen agents that aced benchmarks fail spectacularly in production. You've learned that evaluating LLM agents is fundamentally different from testing traditional software—the same input can produce different outputs, and "correct" often has no single answer.
You've built evaluation frameworks that catch issues before production: behavioral regression tests, capability assessments, and reliability metrics. You understand that the goal isn't 100% test pass rate—it
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-02 | fail→pass | 16,771 | 24,409 | +46% | 1 | 1 | 0% | 2,599 | 4,533 | +74% | 0 | 0 | — |
case-21 | pass→pass | 12,591 | 11,087 | -12% | 1 | 1 | 0% | 1,958 | 2,214 | +13% | 0 | 0 | — |
case-01 | fail→fail | 23,200 | 40,778 | +76% | 1 | 1 | 0% | 4,133 | 6,050 | +46% | 0 | 0 | — |
case-03 | pass→pass | 16,135 | 18,172 | +13% | 1 | 1 | 0% | 2,609 | 2,937 | +13% | 0 | 0 | — |
case-04 | pass→pass | 14,842 | 16,724 | +13% | 1 | 1 | 0% | 2,663 | 3,027 | +14% | 0 | 0 | — |
case-05 | fail→pass | 12,072 | 13,440 | +11% | 1 | 1 | 0% | 2,155 | 2,440 | +13% | 0 | 0 | — |
case-06 | pass→pass | 14,272 | 15,930 | +12% | 1 | 1 | 0% | 2,095 | 2,776 | +33% | 0 | 0 | — |
case-07 | pass→pass | 12,906 | 17,837 | +38% | 1 | 1 | 0% | 2,304 | 3,375 | +46% | 0 | 0 | — |
case-08 | pass→pass | 13,263 | 16,098 | +21% | 1 | 1 | 0% | 2,143 | 2,661 | +24% | 0 | 0 | — |
case-09 | fail→fail | 13,079 | 14,333 | +10% | 1 | 1 | 0% | 2,227 | 2,727 | +22% | 0 | 0 | — |
case-10 | pass→pass | 15,823 | 14,489 | -8% | 1 | 1 | 0% | 2,435 | 2,396 | -2% | 0 | 0 | — |
case-11 | pass→pass | 13,063 | 13,872 | +6% | 1 | 1 | 0% | 2,246 | 2,744 | +22% | 0 | 0 | — |
case-12 | fail→pass | 14,054 | 14,432 | +3% | 1 | 1 | 0% | 1,866 | 2,743 | +47% | 0 | 0 | — |
case-13 | fail→fail | 12,802 | 11,361 | -11% | 1 | 1 | 0% | 2,434 | 2,227 | -9% | 0 | 0 | — |
case-14 | fail→pass | 17,772 | 14,003 | -21% | 1 | 1 | 0% | 2,538 | 2,687 | +6% | 0 | 0 | — |
case-15 | fail→pass | 17,409 | 16,810 | -3% | 1 | 1 | 0% | 2,830 | 3,292 | +16% | 0 | 0 | — |
case-16 | fail→fail | 15,645 | 18,238 | +17% | 1 | 1 | 0% | 2,408 | 3,513 | +46% | 0 | 0 | — |
case-17 | pass→pass | 13,857 | 14,410 | +4% | 1 | 1 | 0% | 2,055 | 2,744 | +34% | 0 | 0 | — |
case-18 | pass→pass | 12,226 | 13,251 | +8% | 1 | 1 | 0% | 2,145 | 2,371 | +11% | 0 | 0 | — |
case-19 | fail→pass | 17,608 | 18,460 | +5% | 1 | 1 | 0% | 2,659 | 3,273 | +23% | 0 | 0 | — |
case-20 | pass→pass | 20,291 | 14,245 | -30% | 1 | 1 | 0% | 3,293 | 2,629 | -20% | 0 | 0 | — |
case-22 | pass→fail | 9,020 | 14,081 | +56% | 1 | 1 | 0% | 1,532 | 2,618 | +71% | 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 +23 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.