Chain-of-thought (CoT) prompting elicits step-by-step reasoning before final answers, dramatically improving accuracy. Mechanism: Ask model to "think step by step" or demonstrate reasoning in examples. Model generates intermediate steps that guide toward correct answer. Implementation: Zero-shot ("Let's think step by step"), few-shot (examples showing reasoning), or structured templates. Why it works: Breaks complex problems into manageable steps, reduces reasoning errors, leverages model's training on step-by-step explanations. Best for: Math problems, logic puzzles, multi-hop reasoning, complex analysis, code debugging. Limitations: Longer outputs (cost/latency), can generate plausible but wrong reasoning, small models may not benefit. Variants: Self-consistency (multiple paths, vote on answer), Tree of Thoughts (explore branches), least-to-most (decompose then solve). Emergent ability: Works best in large models (100B+ parameters), limited effect in smaller models. Best practices: Be explicit about step-by-step format, verify reasoning not just answers, combine with self-consistency for important tasks. One of the most practical prompt engineering techniques.