Chain-of-thought prompting asks the model to reason step by step before giving a final answer.
DefinitionWhat it means
Chain-of-Thought prompting instructs or encourages a model to generate intermediate reasoning steps, arithmetic, logic, or a plan, before it commits to a final answer. This can be triggered with an explicit instruction such as think step by step, with worked reasoning examples, or built into a model that reasons by default. The extra tokens give the model room to decompose a hard problem instead of pattern matching straight to an answer.
Why it mattersWhy you should care
Chain-of-thought reliably raises accuracy on multi-step math, logic, and planning tasks, which is exactly the kind of task that breaks a naive single-shot prompt in production. It also produces an inspectable trace that helps debug why an answer was wrong, though teams must decide whether to show that trace to end users or hide it and return only the final answer, since exposed reasoning can leak sensitive intermediate content or simply be verbose.
At a glanceSee it
Self-consistency samples several independent reasoning paths and returns the majority answer — far more robust than trusting a single chain, though correlated paths can still share one blind spot.
Chain-of-thought is a choice, not a default — it pays off on multi-step problems but wastes latency on simple ones, and a fluent-looking chain can still land on a wrong answer.
Where you see itIn the wild
- Math and logic benchmarks where step-by-step prompting is compared against direct answering.
- Reasoning models that emit a hidden or visible thinking trace before the response.
- Agent debugging logs used to trace why a tool call or decision went wrong.