Chain-of-thought
Also known as: CoT
What is Chain-of-thought?
A prompting technique that asks a model to work through a problem step by step instead of jumping straight to an answer. Spelling out the intermediate steps raises accuracy on maths, logic and multi-step tasks, and shows its working. The same idea, trained into a model, is what powers today's reasoning models.
Why it matters
Asking a model to reason step by step buys two things: better answers on problems that need several stages, and a visible trace you can inspect when something goes wrong. The second is easy to over-trust. The written steps read like the model’s actual reasoning, but they are text generated to look plausible, and a model can reach the right answer through a wrong-looking explanation or the reverse. Treat the chain as a debugging aid and a quality lever, not as proof of how the machine really got there. It also is not free: every intermediate step is more tokens generated, so slower and more expensive, and you spend it where accuracy matters and skip it on simple lookups.
In practice
A finance team uses step-by-step prompting to have a model check calculations in supplier contracts, and accuracy on multi-part clauses climbs sharply. Reading the intermediate steps also lets a reviewer spot exactly where a wrong figure crept in. For simple field extraction they drop the technique, since the added reasoning only slows things down and runs up the token bill for no gain.
