Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Stress-test the winning candidate using Pre-mortem, Red Teaming, and Failure Mode Analysis to expose hidden weaknesses before commitment.
.claude/skills/yogsoth-ai-winner-stress-testing/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-15 | ✗→✓ | ▲ Improved | 95% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 18% | 0% |
| case-01 | ✗→✓ | ▲ Improved | -2% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 32% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 169% | 0% |
Purpose: Subject the convergence winner to rigorous adversarial pressure, identifying failure modes, hidden assumptions, and boundary conditions that could cause the decision to fail in practice.
When to use:
| Metric | Minimum | |--------|---------| | Attack angles | >= 3 distinct failure vectors | | Assumptions challenged | >= 5 | | Pre-mortem scenarios | >= 3 | | Severity threshold | All HIGH severity findings must be addressed |
yamlwinner: <candidate> attack_vectors_applied: [] assumptions_found: [] failure_modes: [] severity_ratings: {} verdict: null # ACCEPT | REJECT | REVISE conditions_for_acceptance: []
| Tactic | When to Deploy | |--------|---------------| | assumption-excavation | Default — extract and challenge winner's assumptions | | adversarial-debate-protocol | When specific weaknesses need formal debate | | multi-perspective-attack | When winner affects multiple stakeholder groups |
yamlstrategy: winner-stress-testing winner: <candidate> assumptions_found: <count> critical_assumptions: <count> failure_modes: - mode: <description> severity: HIGH | MEDIUM | LOW mitigable: true | false verdict: ACCEPT | REJECT | REVISE conditions: [] recommended_modifications: []
<!-- BEGIN available-tables (generated) -->
Optional, no fixed order; the final leaf is always a sop.
| Tactic | When to use | | --- | --- | | adversarial-debate-protocol | Structured debate protocol that constructs an advocate, deploys critic attacks, and renders a judge verdict through iterative rounds. | | assumption-excavation | Systematic extraction, challenge, and sensitivity analysis of assumptions underlying a decision to identify load-bearing beliefs. | | multi-perspective-attack | Assign distinct perspectives to attack a decision from multiple angles, then synthesize findings into a unified assessment. |
<!-- 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-15 | fail→pass | 20,842 | 34,984 | +68% | 1 | 1 | 0% | 2,897 | 5,647 | +95% | 0 | 0 | — |
case-05 | fail→pass | 33,241 | 35,312 | +6% | 1 | 1 | 0% | 4,368 | 5,157 | +18% | 0 | 0 | — |
case-01 | fail→pass | 42,204 | 32,058 | -24% | 1 | 1 | 0% | 5,126 | 5,029 | -2% | 0 | 0 | — |
case-02 | fail→pass | 35,502 | 72,625 | +105% | 1 | 1 | 0% | 4,747 | 6,282 | +32% | 0 | 0 | — |
case-03 | fail→pass | 16,035 | 20,768 | +30% | 1 | 1 | 0% | 1,242 | 3,337 | +169% | 0 | 0 | — |
case-04 | fail→pass | 36,648 | 19,118 | -48% | 1 | 1 | 0% | 4,791 | 3,803 | -21% | 0 | 0 | — |
case-06 | fail→fail | 33,916 | 23,666 | -30% | 1 | 1 | 0% | 4,341 | 4,525 | +4% | 0 | 0 | — |
case-07 | fail→pass | 18,069 | 14,506 | -20% | 1 | 1 | 0% | 2,093 | 3,128 | +49% | 0 | 0 | — |
case-08 | fail→pass | 18,564 | 31,592 | +70% | 1 | 1 | 0% | 2,956 | 4,829 | +63% | 0 | 0 | — |
case-09 | fail→pass | 12,279 | 25,208 | +105% | 1 | 1 | 0% | 1,975 | 3,829 | +94% | 0 | 0 | — |
case-10 | pass→pass | 25,836 | 25,215 | -2% | 1 | 1 | 0% | 3,167 | 3,941 | +24% | 0 | 0 | — |
case-11 | fail→pass | 23,288 | 30,550 | +31% | 1 | 1 | 0% | 3,629 | 4,734 | +30% | 0 | 0 | — |
case-12 | pass→pass | 24,995 | 26,553 | +6% | 1 | 1 | 0% | 3,164 | 4,307 | +36% | 0 | 0 | — |
case-13 | fail→pass | 28,304 | 40,400 | +43% | 1 | 1 | 0% | 4,200 | 6,423 | +53% | 0 | 0 | — |
case-14 | fail→pass | 25,040 | 35,833 | +43% | 1 | 1 | 0% | 3,959 | 5,394 | +36% | 0 | 0 | — |
case-16 | fail→pass | 23,499 | 31,713 | +35% | 1 | 1 | 0% | 2,511 | 4,984 | +98% | 0 | 0 | — |
case-17 | pass→pass | 17,789 | 13,315 | -25% | 1 | 1 | 0% | 2,187 | 2,034 | -7% | 0 | 0 | — |
case-18 | fail→pass | 17,469 | 22,557 | +29% | 1 | 1 | 0% | 2,911 | 4,514 | +55% | 0 | 0 | — |
case-19 | fail→pass | 23,983 | 32,428 | +35% | 1 | 1 | 0% | 3,057 | 4,673 | +53% | 0 | 0 | — |
case-20 | fail→pass | 15,505 | 33,175 | +114% | 1 | 1 | 0% | 1,413 | 5,671 | +301% | 0 | 0 | — |
case-21 | pass→fail | 28,474 | 48,491 | +70% | 1 | 1 | 0% | 4,015 | 8,878 | +121% | 0 | 0 | — |
case-22 | pass→pass | 29,128 | 23,035 | -21% | 1 | 1 | 0% | 3,296 | 3,707 | +12% | 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 +68 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.