How do I write a prompt that forces an AI to prove its answer?

A prompt you can paste

Here is the shape of a proving prompt. Adapt the wording, keep the demands:

"Answer the question. Then, for every claim you made, do three things: label it as a documented fact, a supported inference, or a hypothesis; give me a source I can open and read; and tell me what evidence would make the claim wrong. Do not present a hypothesis as a fact. If you cannot source a claim, say so plainly."

That single instruction does most of the work, because it makes the model separate what it knows from what it is guessing, in the open.

The second half: a fresh model

Proving is not just self-reporting. A model checking its own work grades itself an A every time. So take the answer and paste it into a different model with this:

"Here is an answer from another AI. Find the single weakest claim in it, explain exactly why it is weak, and say what evidence would settle it."

A fresh model with no stake in the answer will flag a hallucination on sight, often with the reason attached.

This is the method, compressed

Forcing proof from one model and cross-examining with another is the core of the Council Method and the Claim Ledger tool. Run it across several models and you have full multi-model verification. The complete set of prompts is free in the Council Protocol, and the reasoning behind them is in Let the AI Be Smart.