Chain of Thought (CoT)

Chain of Thought prompting dramatically improves accuracy on tasks requiring logic, math, or multi-step reasoning. Instead of asking "What is 23 * 17?", you prompt: "Think step by step. What is 23 * 17?" The model then shows its work: "23 * 17 = 23 * 10 + 23 * 7 = 230 + 161 = 391." This simple technique can turn wrong answers into correct ones. Modern models like GPT-5.2-pro and Claude Opus 4.6 have built-in reasoning modes that automatically use chain-of-thought. GPT-5.2-pro offers reasoning effort levels (low/medium/high), and Claude Opus 4.6 has adaptive thinking. These handle CoT internally, so you do not always need to prompt for it explicitly. In production, CoT is most valuable for: classification with explanations (the model explains why it chose a category), complex data analysis (step-by-step calculations), and debugging (walking through code logic). The tradeoff is more output tokens, which means higher cost and latency.