Why 'Think Step by Step' Works

**Chain-of-thought (CoT) prompting**: instructing an AI model to write out intermediate reasoning steps before giving its final answer. **Reasoning chain**: the visible sequence of steps the model produces on the way to that answer. Here's the mechanical part. A language model generates its response one token at a time, and everything it has already written becomes context for what it writes next. A direct question asks the model to make one giant leap — question straight to answer, all the intermediate computation compressed into a single prediction. CoT breaks that leap into small hops. Each hop is an easy prediction, and each written hop anchors the next one. The multi-step structure the problem *needed* now exists on the page instead of nowhere. That's also why CoT shines on math, logic, planning, and multi-constraint analysis, and does little for simple lookups. If there's no intermediate work to externalize, scratch paper is just paper. Zero-shot CoT is the lazy (complimentary) version: one trigger phrase, no demonstrations. It pairs beautifully with few-shot examples, where you show the model a worked reasoning chain to imitate — that pairing is exactly where the next module goes.