Chain-of-thought

Chain-of-thought (often shortened to CoT) is a prompting technique that nudges a model to lay out its reasoning in the open — one step at a time — instead of leaping straight to a verdict. A simple cue like "let's think step by step" is enough to switch many models from terse guessing into a more deliberate mode where they decompose the question, handle each part, and assemble the pieces into a final answer.

The mechanism is intuitive once you see it. A single answer token has very little space to do hard work; a paragraph of reasoning gives the model room to carry intermediate values forward, test a sub-result, and catch its own slips before they reach the conclusion. This is why chain-of-thought lifts performance on arithmetic, word problems, and logic puzzles — and why it is the backbone of richer agent patterns. The ReAct pattern, for instance, is chain-of-thought with tool actions slotted between the reasoning steps, and crafting the cues that trigger good reasoning is a core skill in prompt engineering.

Concrete example: ask "A shop sells pens at 3 for $2; how much for 12 pens?" Without CoT a model might blurt a wrong figure. Prompted to reason, it writes "12 pens is 4 groups of 3; each group costs $2; 4 × $2 = $8," and lands on $8. The visible chain makes the math trustworthy and easy to audit. Because the technique works by shaping how a large language model spends its compute, it costs nothing beyond a few extra tokens.