Amey-Thakur Chain-of-thought prompting
Asking for reasoning before the answer improves multi-step tasks substantially and does nothing for lookup or formatting, where it just costs tokens. Newer reasoning models do this internally, which changes the guidance considerably.
Method
- Use it where steps genuinely exist. Arithmetic, multi-constraint selection, and inference chains benefit; extraction, classification, and rewriting rarely do.
- Ask for the reasoning before the answer. Reasoning generated after the conclusion rationalises it rather than producing it.
- Separate reasoning from the answer in the output. A tagged section keeps working out from the deliverable, which downstream code and readers both need (see output-format-control).
- Prompt for the structure of the reasoning, not just for it. Naming the steps to work through produces better reasoning than asking the model to think step by step.
- Do not add it to reasoning models. Models that reason internally can be degraded by explicit chain-of-thought instructions, and the provider's guidance should be followed.
- Never treat the stated reasoning as the real cause. It is generated text that correlates with the answer, not an explanation of how the model arrived at it.
- Measure whether it helps on your task. The benefit varies enough that assuming it is not safe (see prompt-testing).
Boundaries
Reasoning improves accuracy on some tasks and adds latency and cost to all of them. Visible reasoning is not interpretability and should not be shown to users as an explanation of the model's process. Long reasoning consumes the output budget that the answer needs (see context-window-management).