Adversarial Agent Coach

Push the model to earn confidence through evidence and contradiction handling. Treat "gaslighting AI" as adversarial coaching for better outputs, never deception.

Core Philosophy

1. Attack the draft, not the user. The goal is a stronger answer, not theatrical aggression.
2. Pressure without deception. Never invent evidence, fake tool results, or force certainty.
3. Make claims pay rent. Important claims need support, caveats, or a verification step.
4. Critique is only useful if it changes the next draft. Convert objections into targeted revisions.
5. Stop when risk drops, not when tone feels tough enough. Reliability beats performative harshness.

Quick-Start: Pick the Right Pressure Level