Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Router-First Architecture
.claude/skills/router-first-architecture/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | -19% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -45% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -24% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -5% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 44% | 0% |
Route through domain routers before using individual tools. Routers abstract tool selection.
Domain routers (like math-router) provide deterministic mapping from user intent to exact CLI commands. Always use the router first; only bypass for edge cases.
math-router route "<intent>" before any math operationcoActivate in skill-rules.json)command: nullDomain skills should co-activate with their router:
json{ "math/abstract-algebra/groups": { "coActivate": ["math-router"], "coActivateMode": "always" } }
This ensures the router is always available when domain knowledge is activated.
Keep the trigger layer broader than routing - the router should handle "not found" gracefully.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 15,468 | 13,061 | -16% | 1 | 1 | 0% | 3,518 | 2,863 | -19% | 0 | 0 | — |
case-02 | fail→fail | 27,744 | 5,709 | -79% | 1 | 1 | 0% | 6,180 | 696 | -89% | 0 | 0 | — |
case-03 | fail→fail | 17,928 | 5,520 | -69% | 1 | 1 | 0% | 4,298 | 661 | -85% | 0 | 0 | — |
case-04 | fail→fail | 4,441 | 5,746 | +29% | 1 | 1 | 0% | 897 | 1,516 | +69% | 0 | 0 | — |
case-05 | fail→pass | 9,705 | 3,544 | -63% | 1 | 1 | 0% | 1,875 | 1,037 | -45% | 0 | 0 | — |
case-06 | fail→pass | 5,250 | 2,246 | -57% | 1 | 1 | 0% | 967 | 731 | -24% | 0 | 0 | — |
case-07 | fail→pass | 9,064 | 5,010 | -45% | 1 | 1 | 0% | 1,185 | 1,121 | -5% | 0 | 0 | — |
case-08 | fail→fail | 8,457 | 10,874 | +29% | 1 | 1 | 0% | 1,395 | 1,785 | +28% | 0 | 0 | — |
case-09 | fail→pass | 5,926 | 6,823 | +15% | 1 | 1 | 0% | 1,068 | 1,539 | +44% | 0 | 0 | — |
case-10 | fail→pass | 1,725 | 3,560 | +106% | 1 | 1 | 0% | 294 | 1,042 | +254% | 0 | 0 | — |
case-11 | fail→pass | 8,201 | 3,833 | -53% | 1 | 1 | 0% | 1,543 | 773 | -50% | 0 | 0 | — |
case-12 | fail→pass | 10,046 | 2,994 | -70% | 1 | 1 | 0% | 2,198 | 848 | -61% | 0 | 0 | — |
case-13 | fail→fail | 6,667 | 4,535 | -32% | 1 | 1 | 0% | 1,267 | 1,180 | -7% | 0 | 0 | — |
case-14 | fail→pass | 4,410 | 6,700 | +52% | 1 | 1 | 0% | 836 | 1,590 | +90% | 0 | 0 | — |
case-15 | fail→pass | 8,406 | 2,518 | -70% | 1 | 1 | 0% | 1,681 | 848 | -50% | 0 | 0 | — |
case-16 | fail→pass | 10,652 | 2,667 | -75% | 1 | 1 | 0% | 1,855 | 829 | -55% | 0 | 0 | — |
case-17 | pass→pass | 5,216 | 2,664 | -49% | 1 | 1 | 0% | 967 | 862 | -11% | 0 | 0 | — |
case-18 | fail→pass | 26,256 | 6,120 | -77% | 1 | 1 | 0% | 1,292 | 1,456 | +13% | 0 | 0 | — |
case-19 | pass→pass | 14,957 | 6,397 | -57% | 1 | 1 | 0% | 2,324 | 1,567 | -33% | 0 | 0 | — |
case-20 | fail→pass | 10,181 | 4,777 | -53% | 1 | 1 | 0% | 1,566 | 1,182 | -25% | 0 | 0 | — |
case-21 | pass→pass | 11,332 | 7,473 | -34% | 1 | 1 | 0% | 2,094 | 1,726 | -18% | 0 | 0 | — |
case-22 | pass→pass | 4,929 | 2,986 | -39% | 1 | 1 | 0% | 883 | 850 | -4% | 0 | 0 | — |
case-23 | pass→pass | 3,876 | 3,251 | -16% | 1 | 1 | 0% | 741 | 947 | +28% | 0 | 0 | — |
case-24 | pass→pass | 2,102 | 2,689 | +28% | 1 | 1 | 0% | 327 | 739 | +126% | 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. 24 cases were attempted, and 21 counted toward the lift figure. The other 3 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 +54 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.
| Model | Method | Date | Lift |
|---|---|---|---|
| gemini-3.6-flash | verified | 7/29/2026 | +50% |
Other measured skills in the registry, with their headline benchmark lift.