Chain-of-thought prompting

It helps where a quick answer would skip a step, such as a discount calculation or a decision with several conditions. It also gives you something to audit, because a wrong step is easier to spot than a wrong final number.

How it works

Wei and colleagues showed in 2022 that putting worked examples with step-by-step reasoning into the prompt made large models much better at math word problems and other multi-step reasoning tasks. Later that year Kojima and colleagues found a version that needs no examples: adding "Let's think step by step" before the answer improved results on its own.

Reasoning models now do this internally, so with them you rarely need to ask. With standard models the instruction still helps on multi-step work. Ask for the reasoning in one section and the final answer in another, so an automation can read only the answer.

Example

An operations analyst asks how many extra warehouse staff a holiday peak will need, given the order forecast and the site's shift pattern. She adds "work through this step by step, then give the final number on its own line." The working shows the model assumed eight-hour shifts where the site runs ten, so she corrects that one input and reruns it.

Common mistake

People paste "think step by step" into every prompt. On simple tasks like classification it makes replies slower and longer with no gain in accuracy. Use it where the task has several dependent steps, and test with and without it on your own examples.