Chain-of-thought (CoT)

Chain-of-thought (CoT) is a prompting technique that asks a model to produce intermediate reasoning before its final answer.

Chain-of-thought prompting asks a model to write intermediate steps as it works toward an answer. Research published in 2022 popularized the technique on models that often benefited from explicit step-by-step examples. Current reasoning models handle much of that work internally. In July 2026, Claude Fable 5 uses always-on adaptive thinking, while GPT-5.6 Sol controls internal reasoning through effort settings and modes. For a scheduling problem, ask for a schedule, a list of the constraints it claims to satisfy, and a final verification against the source constraints. That request exposes claims a validator can test without trying to extract the model's private reasoning trace.

Builder example

Manual step-by-step prompting can add latency, produce a persuasive but wrong explanation, or reduce performance on models already trained to reason internally. A deployed workflow should start with a direct task, explicit success criteria, and a way to verify the answer. Add structured decomposition only when an evaluation set exposes a specific failure that decomposition repairs.

A model proposes a meeting schedule, but the final table does not show whether every room and time constraint was satisfied.

Ask for the schedule, a concise constraint check, and a final verification. Add an explicit plan only when representative tests show that decomposition fixes a recurring miss.

Common confusion: A concise rationale, a list of checked constraints, and a hidden reasoning trace are different outputs. Providers may keep internal reasoning private. Any visible explanation still needs independent checks because a coherent explanation can accompany a wrong answer.