Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Strategy: Classic triangular debate — Critic attacks, Defender responds, Judge adjudicates. Based on Irving AI Safety via Debate with Toulmin argumentation structure.
.claude/skills/yogsoth-ai-critic-defender-judge/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-03 | ✗→✓ | ▲ Improved | 113% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 14% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 25% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -8% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -9% | 0% |
Classic adversarial triangle: one agent attacks, one defends, one judges.
| Parameter | S | M | L | |---|---|---|---| | Debate rounds | 4 | 8 | 12 | | Participating agents | 3 | 5 | 8 | | Coverage dimensions | 3 | 5 | 7 | | External evidence searches | 2 | 5 | 10 |
debate-architect → [for each round]:
debate-critic → debate-defender → debate-judge
→ confidence-calibration → (escalate or terminate)
→ debate-transcript-analysis → verdict-synthesis<!-- BEGIN available-tables (generated) -->
Optional, no fixed order; the final leaf is always a sop.
| Tactic | When to use | | --- | --- | | evidence-tournament | Tactic: Evidence gathering, cross-examination, and quality judgment. External evidence is collected, presented, challenged, and scored for relevance and reliability. | | stress-test-dialectical-escalation | Tactic: Progressive debate escalation based on confidence thresholds. Each round increases attack sophistication until defender collapses or proves resilient. |
Optional, no fixed order; the final leaf is always a sop.
| SOP | When to use | | --- | --- | | confidence-calibration | Calibrates confidence scores based on debate progression. Determines whether to escalate, continue, or terminate based on cumulative evidence. | | debate-architect | Designs debate structure based on artifact type — selects attack vectors, assigns perspectives, determines escalation ladder, and configures round parameters. | | debate-critic | Generates structured criticism from attack stance using Toulmin model. Produces claims, grounds, warrants, and rebuttals targeting artifact weaknesses. | | debate-defender | Responds to attacks with counter-evidence and counter-arguments. Defends artifact using evidence, clarification, and rebuttal while acknowledging valid criticisms. | | debate-judge | Evaluates debate exchanges, adjudicates argument quality, and produces round verdicts with confidence scores and reasoning. | | evidence-scout | Searches for external evidence supporting or opposing specific claims. Returns structured evidence with source assessment and relevance scoring. |
<!-- 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-01 | pass→pass | 26,527 | 34,162 | +29% | 1 | 1 | 0% | 3,864 | 5,939 | +54% | 0 | 0 | — |
case-02 | fail→fail | 19,917 | 24,788 | +24% | 1 | 1 | 0% | 2,812 | 4,457 | +58% | 0 | 0 | — |
case-03 | fail→pass | 23,907 | 27,794 | +16% | 1 | 1 | 0% | 2,444 | 5,210 | +113% | 0 | 0 | — |
case-04 | pass→pass | 22,186 | 23,085 | +4% | 1 | 1 | 0% | 4,330 | 5,374 | +24% | 0 | 0 | — |
case-05 | pass→fail | 15,025 | 34,063 | +127% | 1 | 1 | 0% | 2,706 | 6,738 | +149% | 0 | 0 | — |
case-06 | pass→fail | 19,006 | 24,584 | +29% | 1 | 1 | 0% | 3,537 | 4,931 | +39% | 0 | 0 | — |
case-07 | fail→pass | 15,704 | 13,690 | -13% | 1 | 1 | 0% | 2,053 | 2,343 | +14% | 0 | 0 | — |
case-08 | fail→pass | 23,427 | 22,881 | -2% | 1 | 1 | 0% | 3,532 | 4,410 | +25% | 0 | 0 | — |
case-09 | fail→pass | 29,248 | 25,356 | -13% | 1 | 1 | 0% | 5,212 | 4,775 | -8% | 0 | 0 | — |
case-10 | pass→pass | 17,046 | 12,763 | -25% | 1 | 1 | 0% | 2,854 | 2,720 | -5% | 0 | 0 | — |
case-11 | pass→pass | 21,965 | 13,126 | -40% | 1 | 1 | 0% | 3,462 | 2,745 | -21% | 0 | 0 | — |
case-12 | fail→pass | 18,662 | 15,381 | -18% | 1 | 1 | 0% | 3,482 | 3,168 | -9% | 0 | 0 | — |
case-13 | fail→pass | 12,309 | 5,339 | -57% | 1 | 1 | 0% | 2,078 | 1,579 | -24% | 0 | 0 | — |
case-14 | pass→pass | 19,799 | 25,179 | +27% | 1 | 1 | 0% | 2,993 | 4,502 | +50% | 0 | 0 | — |
case-15 | pass→pass | 14,961 | 31,309 | +109% | 1 | 1 | 0% | 2,271 | 4,530 | +99% | 0 | 0 | — |
case-16 | fail→pass | 17,399 | 14,214 | -18% | 1 | 1 | 0% | 2,567 | 2,979 | +16% | 0 | 0 | — |
case-17 | fail→pass | 17,968 | 11,273 | -37% | 1 | 1 | 0% | 1,742 | 2,590 | +49% | 0 | 0 | — |
case-18 | pass→fail | 14,968 | 3,898 | -74% | 1 | 1 | 0% | 2,346 | 1,257 | -46% | 0 | 0 | — |
case-19 | pass→pass | 20,118 | 14,276 | -29% | 1 | 1 | 0% | 3,246 | 3,044 | -6% | 0 | 0 | — |
case-20 | fail→pass | 10,014 | 3,421 | -66% | 1 | 1 | 0% | 1,579 | 1,185 | -25% | 0 | 0 | — |
case-21 | fail→pass | 15,620 | 5,162 | -67% | 1 | 1 | 0% | 1,338 | 1,411 | +5% | 0 | 0 | — |
case-22 | fail→pass | 23,069 | 4,124 | -82% | 1 | 1 | 0% | 1,323 | 1,358 | +3% | 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 +36 percentage points is the difference between those two pass rates over the 22 comparable cases. 3 cases got worse with the skill loaded, and they are 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.