2.4 — Rule #3 — Get a verifiable answer

🟢 In plain words — do not merely ask the model to “think harder”. State the goal, authorised data, constraints and criteria you will use to check the answer.

Chain-of-Thought refers to a model’s intermediate reasoning. Earlier prompting research showed that intermediate steps could help some models with arithmetic or logic. With current reasoning models, you do not need to request their full private reasoning.

Left to right: you guide the model more and more.

A practical method

1. Define the expected result.
2. Supply authorised data or sources.
3. State important constraints.
4. Request the final answer, a concise rationale and uncertainties.
5. For calculations, request the operations needed for verification.
6. Test on several cases and use calculation or search tools when appropriate.

Caution

Reasoning models already reason internally. Asking them to “think step by step” can be unnecessary or counterproductive. Control the observable inputs, sources, tool calls, outputs and success criteria.

Primary source: OpenAI reasoning best practices.

✅ Summary

• Use goal + data + constraints + success criteria.
• Ask for a concise, verifiable rationale, not full private reasoning.
• Measure quality on outputs with tools and a fixed test set.