Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Complexity-based task routing with Q-Learning optimization, Agent Booster WASM fast-path, and Mixture-of-Experts model selection.
.claude/skills/a5c-ai-smart-routing/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | 17% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 272% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -31% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -57% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -68% | 0% |
| Tier | Target | Latency | Cost | |------|--------|---------|------| | Agent Booster | Simple transforms (var-to-const, add-types) | <1ms | $0 | | Medium | Standard coding tasks | ~500ms | Low | | Complex | Multi-agent swarm coordination | 2-5s | Higher |
var-to-const - Variable declaration modernizationadd-types - TypeScript type annotation insertionadd-error-handling - Try/catch wrapper insertionasync-await - Promise chain to async/await conversionextract-function - Code block extraction to named functionsadd-jsdoc - Documentation generationagents/optimizer/ - Performance and cost optimizationagents/architect/ - Complex task decompositionInvoke via babysitter process: methodologies/ruflo/ruflo-task-routing
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 3,700 | 5,999 | +62% | 1 | 1 | 0% | 632 | 1,488 | +135% | 0 | 0 | — |
case-02 | fail→pass | 14,588 | 16,101 | +10% | 1 | 1 | 0% | 2,978 | 3,489 | +17% | 0 | 0 | — |
case-03 | fail→fail | 14,385 | 8,751 | -39% | 1 | 1 | 0% | 3,373 | 2,225 | -34% | 0 | 0 | — |
case-04 | fail→pass | 3,734 | 11,187 | +200% | 1 | 1 | 0% | 587 | 2,183 | +272% | 0 | 0 | — |
case-05 | fail→fail | 14,034 | 9,163 | -35% | 1 | 1 | 0% | 3,092 | 2,284 | -26% | 0 | 0 | — |
case-06 | fail→fail | 10,364 | 9,008 | -13% | 1 | 1 | 0% | 2,515 | 1,715 | -32% | 0 | 0 | — |
case-07 | fail→pass | 14,608 | 8,400 | -42% | 1 | 1 | 0% | 2,557 | 1,760 | -31% | 0 | 0 | — |
case-08 | fail→pass | 12,989 | 3,547 | -73% | 1 | 1 | 0% | 2,283 | 985 | -57% | 0 | 0 | — |
case-09 | fail→pass | 11,952 | 2,251 | -81% | 1 | 1 | 0% | 1,895 | 610 | -68% | 0 | 0 | — |
case-10 | fail→pass | 8,839 | 1,254 | -86% | 1 | 1 | 0% | 1,558 | 442 | -72% | 0 | 0 | — |
case-11 | fail→pass | 13,816 | 1,780 | -87% | 1 | 1 | 0% | 2,405 | 561 | -77% | 0 | 0 | — |
case-12 | fail→fail | 5,863 | 1,625 | -72% | 1 | 1 | 0% | 1,063 | 557 | -48% | 0 | 0 | — |
case-13 | pass→pass | 10,616 | 1,661 | -84% | 1 | 1 | 0% | 1,827 | 548 | -70% | 0 | 0 | — |
case-14 | pass→pass | 9,724 | 1,735 | -82% | 1 | 1 | 0% | 1,667 | 568 | -66% | 0 | 0 | — |
case-15 | fail→fail | 7,946 | 1,394 | -82% | 1 | 1 | 0% | 1,570 | 441 | -72% | 0 | 0 | — |
case-16 | pass→pass | 9,983 | 2,193 | -78% | 1 | 1 | 0% | 1,744 | 651 | -63% | 0 | 0 | — |
case-17 | fail→fail | 16,325 | 3,262 | -80% | 1 | 1 | 0% | 1,174 | 455 | -61% | 0 | 0 | — |
case-18 | fail→pass | 8,066 | 1,118 | -86% | 1 | 1 | 0% | 1,370 | 410 | -70% | 0 | 0 | — |
case-19 | fail→pass | 15,171 | 8,753 | -42% | 1 | 1 | 0% | 2,532 | 1,863 | -26% | 0 | 0 | — |
case-20 | fail→pass | 11,832 | 4,366 | -63% | 1 | 1 | 0% | 1,881 | 977 | -48% | 0 | 0 | — |
case-21 | fail→fail | 10,518 | 4,726 | -55% | 1 | 1 | 0% | 1,828 | 1,118 | -39% | 0 | 0 | — |
case-22 | fail→pass | 10,414 | 3,551 | -66% | 1 | 1 | 0% | 1,710 | 872 | -49% | 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 21 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 +50 percentage points is the difference between those two pass rates over the 21 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.