Chain-of-Thought Prompting Trades Latency for Reasoning Depth Chain-of-thought (CoT) prompting improves a model's performance on multi-step problems by instructing it to produce intermediate reasoning steps before committing to a final answer. Rather than jumping straight to an output, the model externalises its working, which appears to help most on tasks with sequential structure — arithmetic, symbolic manipulation, and logic. The mechanism is prompt-level, not architectural: no weights change, only the shape of the requested output. This makes CoT a cheap first lever to try before reaching for fine-tuning or tool use. The trade-off is token cost and latency, since every intermediate step is generated text. It works best on problems that genuinely decompose into steps; on single-hop factual recall it adds cost without benefit. Why this matters: it's the lowest-effort intervention that reliably lifts reasoning-heavy task performance, so it's the default thing to test first. Links: new standalone note Tags: prompt-engineering, chain-of-thought, reasoning, inference-cost, prompting-techniques Review flags: none