Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Identifies agreement and disagreement patterns across multiple perspective evaluations. Maps consensus clusters and persistent divergence points.
.claude/skills/yogsoth-ai-divergence-detection/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-05 | ✗→✓ | ▲ Improved | -38% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -56% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -37% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -24% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -55% | 0% |
Identifies agreement/disagreement across perspectives.
Subagent — spawned via subagent-spawning/spawn-agent.
Divergence analysis requires comparing all perspective outputs simultaneously in dedicated context without being anchored to any single perspective.
One unit = one divergence analysis per round.
<!-- BEGIN available-tables (generated) -->
Optional, no fixed order; the final leaf is always a sop.
| SOP | When to use | | --- | --- | | spawn-agent | Spawn a customized CC subagent with full MCP tool access. Used by SOPs that declare execution: subagent. |
<!-- 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-22 | fail→fail | 5,066 | 14,525 | +187% | 1 | 1 | 0% | 497 | 2,773 | +458% | 0 | 0 | — |
case-01 | fail→fail | 5,629 | 6,161 | +9% | 1 | 1 | 0% | 936 | 1,305 | +39% | 0 | 0 | — |
case-02 | fail→fail | 18,710 | 5,766 | -69% | 1 | 1 | 0% | 2,682 | 1,223 | -54% | 0 | 0 | — |
case-03 | pass→pass | 7,533 | 4,882 | -35% | 1 | 1 | 0% | 1,312 | 1,134 | -14% | 0 | 0 | — |
case-04 | pass→pass | 6,125 | 2,612 | -57% | 1 | 1 | 0% | 1,030 | 737 | -28% | 0 | 0 | — |
case-05 | fail→pass | 15,535 | 6,564 | -58% | 1 | 1 | 0% | 2,328 | 1,450 | -38% | 0 | 0 | — |
case-06 | fail→pass | 10,057 | 2,616 | -74% | 1 | 1 | 0% | 1,616 | 710 | -56% | 0 | 0 | — |
case-07 | fail→pass | 6,882 | 2,347 | -66% | 1 | 1 | 0% | 980 | 616 | -37% | 0 | 0 | — |
case-08 | fail→pass | 7,615 | 4,104 | -46% | 1 | 1 | 0% | 1,110 | 844 | -24% | 0 | 0 | — |
case-09 | pass→pass | 7,231 | 3,444 | -52% | 1 | 1 | 0% | 1,257 | 861 | -32% | 0 | 0 | — |
case-10 | fail→pass | 11,930 | 3,025 | -75% | 1 | 1 | 0% | 1,746 | 778 | -55% | 0 | 0 | — |
case-11 | fail→pass | 10,058 | 6,931 | -31% | 1 | 1 | 0% | 1,444 | 1,381 | -4% | 0 | 0 | — |
case-12 | fail→pass | 12,857 | 1,877 | -85% | 1 | 1 | 0% | 1,993 | 541 | -73% | 0 | 0 | — |
case-13 | fail→pass | 14,370 | 2,266 | -84% | 1 | 1 | 0% | 2,584 | 630 | -76% | 0 | 0 | — |
case-14 | pass→pass | 11,683 | 5,633 | -52% | 1 | 1 | 0% | 1,796 | 1,124 | -37% | 0 | 0 | — |
case-15 | fail→pass | 6,273 | 2,105 | -66% | 1 | 1 | 0% | 885 | 581 | -34% | 0 | 0 | — |
case-16 | fail→pass | 8,648 | 2,069 | -76% | 1 | 1 | 0% | 1,434 | 578 | -60% | 0 | 0 | — |
case-17 | fail→pass | 9,266 | 6,498 | -30% | 1 | 1 | 0% | 1,385 | 1,347 | -3% | 0 | 0 | — |
case-18 | fail→pass | 9,474 | 1,657 | -83% | 1 | 1 | 0% | 1,512 | 529 | -65% | 0 | 0 | — |
case-19 | fail→pass | 21,926 | 3,085 | -86% | 1 | 1 | 0% | 966 | 794 | -18% | 0 | 0 | — |
case-20 | pass→pass | 14,662 | 20,051 | +37% | 1 | 1 | 0% | 2,247 | 3,392 | +51% | 0 | 0 | — |
case-21 | fail→fail | 4,304 | 13,832 | +221% | 1 | 1 | 0% | 719 | 3,076 | +328% | 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 +59 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.