What is chain-of-thought prompting?

Chain-of-thought prompting is an instruction about order: work through the problem in steps, then give the answer. "Think step by step before answering" is the whole technique in one sentence. It helps because each step the model writes becomes part of what it reads while producing the next one, so the answer is built on stated intermediate results instead of arrived at in a single jump.

Two things follow. The reasoning is visible, so a wrong answer can be traced to the line that went wrong — often a misreading of the question rather than a slip in the arithmetic. And it costs tokens and latency, which is why it belongs on multi-step work such as calculation, logic, or weighing something against several criteria, not on every prompt you write.

One wrinkle is current. Recent Claude models reason adaptively and will often do this unasked on a hard question. Asking explicitly still matters when thinking is off, and when you want the steps shown to you rather than kept internal.