| name | counterexample-search |
| description | Before asserting a universal claim — always, never, all, none, guaranteed, impossible — spend one bounded search for a counterexample, starting at the edges of the quantifier's domain. |
| license | MIT |
counterexample-search
Trigger (observable): The draft contains a universal quantifier or impossibility claim ('this always terminates', 'no user can reach this state', 'works for all inputs'), or the task asks whether a property holds in general.
When NOT to activate: Claims already scoped to observations ('in the 30 cases tested...'); definitional truths (tautologies, type guarantees enforced by a compiler); casual prose where the quantifier is rhetorical and nothing depends on it.
Procedure
- Extract each universal claim from the draft verbatim.
- For each, describe the domain the quantifier ranges over, and go to its edges: empty, zero, one, duplicates, maximum, negative, concurrent, malformed, adversarial.
- Construct the single most promising counterexample candidate — the edge case with the best chance of breaking the claim — and actually test or trace it, not just contemplate it.
- If a counterexample is found: weaken the claim to its true scope ('holds for non-empty inputs') or fix the artifact so the claim holds.
- If none is found after the bounded attempt: keep the claim but note the strongest case tried, so the search is auditable.
Required output
Per universal claim: the strongest counterexample candidate, the result of actually testing it, and the claim kept / weakened / withdrawn accordingly.
Verification
- For each universal claim, a specific counterexample candidate is named and actually executed or traced — not just contemplated.
- Claims broken by the search are weakened or withdrawn in the final text.
- Net effect: every universal claim in the final answer either survived a named, executed counterexample attempt or was weakened to observed scope.
Known risk: Unbounded hunting — burning the session on exotic cases for low-stakes claims. Mitigation: one strongest candidate per claim, actually tested, then move on.
Max intended cost: ≤200 added output tokens plus at most one execution/trace per universal claim.
Evidence status: DESIGNED — specified from documented reasoning-failure modes; not yet executed as a packaged skill.
Lineage: Derived from a documented reasoning-failure mode — reaching a plausible answer and stopping without probing it — combined with an adversarial-probe verification pattern from the broader Piénsalo skill family.
