Chain of thought CoT

Prompting a model to write out its intermediate steps before the final answer.

Asking a model to work through a problem before answering improves accuracy on tasks with several dependent steps, because each step is generated with the previous ones visible in the context rather than compressed into one leap.

It costs tokens and latency, and it does not help on retrieval, formatting or single-step tasks. On a reasoning-model it is redundant - that behaviour is already trained in.

Written steps are also not a reliable explanation of how the answer was produced. Treat them as scratch work that improves the result, not as an audit trail.

Where this comes up

Prompts, configs and tutorials in the library that touch this term.