Definition
Chain-of-thought, or CoT, is a sequence of intermediate reasoning text generated as part of producing an answer. Chain-of-thought prompting is a way to elicit that sequence, for example through worked examples or a request for intermediate steps.
In an autoregressive model, generated reasoning tokens become context for later tokens. Working through intermediate steps can help on tasks where producing a direct answer is less reliable.
Simple example
A prompt shows worked word problems with their calculations. It then asks: “Four boxes each hold six batteries. After removing three batteries, how many remain?” The model writes “4 times 6 = 24; 24 minus 3 = 21” before answering 21. The examples are the prompting method. The generated calculations are the chain-of-thought.
Why it matters
Chain-of-thought prompting can improve performance on some arithmetic, symbolic, planning, and multi-step question-answering tasks. Intermediate text may expose a mistake that would be harder to spot in a bare answer.
Generating explicit intermediate reasoning typically uses more tokens and takes longer. A plausible sequence can still end in a wrong answer. Evaluate the final result rather than treating longer reasoning as better reasoning.
One important nuance
Visible chain-of-thought is generated text, not a record of all the computation behind an answer. It can omit steps or fail to reflect what produced the result. A useful answer may expose no reasoning trace at all. An application should ask for the specific evidence, calculations, or decision factors it needs.