Chain-of-Thought Prompting
Chain-of-thought prompting asks a model to work through its reasoning step by step before giving a final answer. Spelling out the intermediate steps measurably improves accuracy on multi-step problems like math, logic, and planning.
Also known as: CoT prompting, CoT, chain of thought
Ask a model for just the answer to a multi-step problem and it often jumps to a wrong one. Chain-of-thought prompting instead asks it to reason through the steps first — “think step by step” — so it builds the answer on visible intermediate work. The technique comes from a 2022 Google Brain paper (Wei et al., “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models”), which showed that on math, logic, and planning tasks, that single change reliably raises accuracy — because the model isn’t trying to produce the conclusion in one leap.
There are two common ways to trigger it. Few-shot CoT puts a couple of worked examples in the prompt, so the model imitates the reasoning pattern. Zero-shot CoT skips the examples and just appends a cue like “let’s think step by step” — cheaper, and often enough on capable models. A related trick, self-consistency, samples several independent chains for the same question and takes the majority answer, trading extra tokens for a further accuracy bump on hard problems.
Two caveats keep this honest. First, cost and latency: more reasoning means more tokens and slower responses, so CoT is worth it for genuinely hard problems and overkill for simple lookups. Second, the visible chain is not a guaranteed window into how the model actually computed the answer — a plausible-looking rationale can accompany a wrong result, which is why CoT improves accuracy without making outputs self-explaining or fully trustworthy.
Newer reasoning models bake this in, doing extended internal chains of thought before answering rather than needing to be prompted for it — which is why “just think step by step” matters less with them than it did. The shift showed up vividly when DeepSeek used reinforcement learning to auto-generate step-by-step reasoning data at scale, a turn we unpacked on the show. For where this fits among the broader levers you have, see prompt vs. context vs. memory engineering, or the AI reasoning overview for the full arc from prompting to reasoning models.
