Chain-of-thought

Chain-of-thought (CoT) prompting is a technique that instructs Claude to break a complex problem into a sequence of intermediate reasoning steps before producing a final answer. Rather than jumping directly to a conclusion, Claude is guided to 'show its work,' which reduces logical leaps, mathematical errors, and unsupported analytical claims. The core idea is that structured deliberation produces more accurate, nuanced, and verifiable outputs than immediate text generation. There are two broad implementation paths. The simpler path uses natural-language instructions in the prompt itself — phrases like 'think step by step' or 'reason through this in <thinking> tags before answering.' This works with any model and any plan. The more powerful path uses the API's native thinking parameter, which lets Claude allocate dedicated compute to reasoning and returns structured thinking blocks alongside the final response. Anthropic's thinking capability has evolved rapidly. Early versions required setting a fixed token budget (budget_tokens). Newer models — Opus 4.6+, Sonnet 4.6+, and later releases — use Adaptive Thinking, where Claude dynamically decides how much to reason based on query complexity and a developer-chosen effort level (low, medium, or high). The older budget_tokens approach is deprecated on these newer models.