Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Generate counterexamples (monsters), attempt monster-barring, incorporate surviving counterexamples as lemma refinements (Lakatos method).
.claude/skills/yogsoth-ai-counterexample-heuristics/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-03 | ✗→✓ | ▲ Improved | -28% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -45% | 0% |
| case-20 | ✗→✓ | ▲ Improved | -25% | 0% |
| case-22 | ✗→✓ | ▲ Improved | -10% | 0% |
| case-10 | ✓→✗ | ▼ Worse | -29% | 0% |
counterexample-generation to produce candidate monstersa. Dispatch monster-barring-attempt — can it be excluded legitimately? b. If barring succeeds: record as excluded, note narrowed scope c. If barring fails: counterexample is genuine
claim-refinement:<!-- BEGIN available-tables (generated) -->
Optional, no fixed order; the final leaf is always a sop.
| SOP | When to use | | --- | --- | | claim-refinement | Propose a refined claim that survives counterexamples while preserving maximum explanatory power (Lakatos lemma-incorporation). | | counterexample-generation | Systematically generate counterexamples (monsters) to a given claim using diverse heuristic strategies. | | monster-barring-attempt | Attempt to exclude a counterexample as illegitimate by tightening definitions or preconditions (Lakatos monster-barring). |
<!-- 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 | 13,495 | 9,343 | -31% | 1 | 1 | 0% | 2,176 | 1,931 | -11% | 0 | 0 | — |
case-02 | pass→pass | 37,880 | 5,403 | -86% | 1 | 1 | 0% | 2,276 | 1,209 | -47% | 0 | 0 | — |
case-03 | fail→pass | 15,400 | 9,266 | -40% | 1 | 1 | 0% | 2,402 | 1,728 | -28% | 0 | 0 | — |
case-04 | pass→pass | 10,076 | 2,775 | -72% | 1 | 1 | 0% | 1,436 | 786 | -45% | 0 | 0 | — |
case-05 | pass→pass | 4,968 | 2,588 | -48% | 1 | 1 | 0% | 775 | 762 | -2% | 0 | 0 | — |
case-06 | pass→pass | 12,780 | 5,973 | -53% | 1 | 1 | 0% | 1,842 | 1,329 | -28% | 0 | 0 | — |
case-07 | pass→pass | 10,271 | 9,467 | -8% | 1 | 1 | 0% | 2,012 | 1,888 | -6% | 0 | 0 | — |
case-08 | pass→pass | 12,267 | 18,049 | +47% | 1 | 1 | 0% | 2,243 | 3,241 | +44% | 0 | 0 | — |
case-09 | pass→pass | 12,419 | 11,306 | -9% | 1 | 1 | 0% | 2,107 | 2,525 | +20% | 0 | 0 | — |
case-10 | pass→fail | 10,522 | 6,160 | -41% | 1 | 1 | 0% | 1,859 | 1,317 | -29% | 0 | 0 | — |
case-11 | pass→pass | 11,315 | 10,619 | -6% | 1 | 1 | 0% | 1,809 | 1,942 | +7% | 0 | 0 | — |
case-12 | fail→pass | 11,157 | 4,340 | -61% | 1 | 1 | 0% | 1,845 | 1,019 | -45% | 0 | 0 | — |
case-13 | fail→fail | 6,035 | 2,173 | -64% | 1 | 1 | 0% | 890 | 688 | -23% | 0 | 0 | — |
case-14 | fail→fail | 6,756 | 2,182 | -68% | 1 | 1 | 0% | 1,106 | 676 | -39% | 0 | 0 | — |
case-15 | pass→pass | 9,279 | 5,163 | -44% | 1 | 1 | 0% | 1,465 | 1,149 | -22% | 0 | 0 | — |
case-16 | pass→pass | 12,862 | 2,567 | -80% | 1 | 1 | 0% | 2,267 | 818 | -64% | 0 | 0 | — |
case-17 | pass→pass | 10,932 | 6,004 | -45% | 1 | 1 | 0% | 1,897 | 1,382 | -27% | 0 | 0 | — |
case-18 | pass→pass | 8,644 | 2,974 | -66% | 1 | 1 | 0% | 1,486 | 885 | -40% | 0 | 0 | — |
case-19 | pass→pass | 15,788 | 11,205 | -29% | 1 | 1 | 0% | 2,772 | 2,275 | -18% | 0 | 0 | — |
case-20 | fail→pass | 5,161 | 1,954 | -62% | 1 | 1 | 0% | 843 | 635 | -25% | 0 | 0 | — |
case-21 | pass→pass | 9,130 | 2,456 | -73% | 1 | 1 | 0% | 1,538 | 808 | -47% | 0 | 0 | — |
case-22 | fail→pass | 18,089 | 13,912 | -23% | 1 | 1 | 0% | 2,825 | 2,538 | -10% | 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 +14 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.