Red Team Prompt to Break Claude's Agreement Bias

LLMs are trained with RLHF to produce responses users rate highly, which creates a systematic bias toward agreement and affirmation. By explicitly assigning a 'destruction' role before sharing any content, you override the model's default cooperative framing before it has anything to be agreeable about. The specificity of the failure categories — flawed assumptions, overlooked risks, second-order effects, likely failure points — acts as a checklist that forces the model to probe multiple dimensions rather than settling on one or two surface-level objections. Vague instructions like 'be critical' leave room for the model to be mildly skeptical; this prompt closes that loophole. Setting the role constraint ahead of the content matters. If you share the idea first and then ask for criticism, the model has already generated positive framing internally. Front-loading the adversarial role primes the entire response generation toward critique from the start.