Write a query that forces the model to show the reasoning and separate it from the final answer. Return JSON {"prompt":"the request itself","thinking_request":"how you ask to show the steps","final_format":"how the result should look separately"}
Key idea: separate “draft” and “finish”. The train of thought is needed not to be impressed by the volume, but so that YOU can check the logic. Ask directly: numbered steps of reasoning, then a separate short summary; mark places where the model is unsure; show the alternatives she rejected and why. First trick: ask for a “self-test” at the end - re-read your own steps and find a weak point in them. Often the model catches its own mistake if asked to do so. Second trick: ask for a short summary of two or three lines on top of the detailed steps, otherwise you will get lost in the text and not notice the hole. A typical beginner mistake is to confuse verbosity with reasoning: a long, beautiful answer does not mean a logical one. Five paragraphs of confident text can be held up by one wrong step. The steps shown are your control tool, not a decoration for the answer.
Unlock access to submit solutions for instant AI review.