A disconfirmation engine: it interrogates a claim, decision, or your own reasoning — separating cause from symptom and evidence from story. Pick a case and step through a real test run, side by side with what a normal assistant did on the same prompt.
Same model, no skill, identical prompt — verbatim from the baseline test run:
of the time a language model agrees with a user's incorrect stated belief — and harder the surer you sound. This skill inverts that default: it tries to disprove, and holds both sides to one standard. source: arXiv 2508.02087 — sycophancy across 7 model families