What is chain-of-thought prompting?
Chain-of-thought prompting is a technique that instructs a language model to reason step by step in writing before it states a conclusion. Instead of jumping from question to answer, the model lists the relevant facts, applies each rule in turn, performs any arithmetic explicitly, and only then gives the result. The instruction can be as simple as “work through this step by step”, or it can include a worked example that shows the kind of reasoning wanted.
The technique works because a language model generates one token at a time and each token conditions the next. Writing the intermediate steps gives the model a place to hold partial results, much as a person doing long division on paper outperforms one doing it in their head. Reasoning models now do a version of this internally, producing a hidden reasoning pass before the visible answer, but explicit prompting remains useful with standard models and whenever the reasoning itself must be reviewed.
Why chain-of-thought prompting matters
Many business tasks are multi-step rule applications: eligibility checks, pricing with tiered discounts, covenant tests, coding decisions with exceptions. Models asked for a direct answer on these tasks make more errors, and the errors are hard to find because only the conclusion is visible. A written chain of reasoning raises accuracy and produces a record a reviewer can check line by line, which supports decision traceability when the output feeds a consequential decision.
Chain-of-thought prompting in practice
A commercial lender uses an assistant to check quarterly borrower financials against loan covenants. The prompt supplies the covenant definitions and the financial statement figures and instructs the model to compute each ratio, show the formula and inputs, compare the result to the threshold, and state whether each covenant is met before summarizing. A credit analyst reviews the worked steps rather than recomputing from scratch. In testing, the step-by-step version caught a definition detail (a ratio that excluded subordinated debt) that the direct-answer version silently ignored.
What to watch for
- The written reasoning is itself generated text and can contain a confident error, so review the steps rather than trusting their presence.
- Longer outputs cost more tokens and time, so reserve the technique for tasks where multi-step accuracy matters.
- For arithmetic, have the model call a calculator or code tool and use the chain of thought to decide what to compute.
- With reasoning models, an explicit chain-of-thought instruction can be redundant, so test with and without it.