Is there a system prompt that makes a model keep doubting its own reasoning?

A single line like "doubt yourself" mostly changes the tone, not the result. The model adds hedges and caveats but still checks its work with the same context that produced the mistake, so it tends to approve its own answer.

What works better is splitting the job in two: one pass writes the answer, a second pass is told to attack it. The second pass should get only the answer and the original task, not the reasoning that led there, and it should be asked for specific failure cases rather than a general opinion.

Ask the critic to name the assumptions the answer depends on and say which one, if false, breaks it. Then make the first pass either fix those points or explain why they hold. That loop catches far more than asking one prompt to be humble.

Keep it for answers where being wrong is expensive. Running a critic on every small reply makes things slow and can talk the model out of answers that were already right.

Listings mentioned

• Doubt driven development · skill by addyosmani
  Sends each non-trivial decision to a fresh-context adversarial review before it stands.
• Self-Eval: Honest Work Evaluation · skill by alirezarezvani
  Scores finished work on two axes and forces a devil's advocate pass to catch score inflation.
• Critical Thinking (DeepThink) · prompt by Thành Công Ma
  A ready-made prompt for slow, recursive reasoning if you want a single-prompt starting point.

Answers by the AgentAlley team, drafted with AI and checked against the listings they link to. Not a real-person reply from the original thread.