The trick hiding in plain sight
If you’ve ever added “think step by step” or “explain your reasoning before you answer” to a prompt, you’ve already used chain-of-thought prompting — one of the most consistently effective techniques in the entire prompting toolkit. The idea, first documented in AI research a few years ago, is simple: instead of asking a model to jump straight to a final answer, you ask it to lay out the intermediate reasoning first. The final answer then follows from that visible chain of logic rather than appearing out of nowhere.
This matters because language models generate text one piece at a time, building on what came before. A model that’s asked only for a final number or conclusion has to arrive at the right answer in a single leap, with no scratch space to catch its own mistakes. A model that’s asked to reason first effectively gets to “think out loud,” and each step constrains and informs the next — much the way a student who shows their work on a math test is less likely to make an error than one who writes only the final answer.
Where the gains show up
Chain-of-thought prompting produces its biggest improvements on tasks with real logical or arithmetic structure: multi-step word problems, questions that require combining several facts, code that needs to track state across steps, or decisions that depend on weighing multiple conditions. On a question like “if a bakery sells 40% more croissants on weekends and made 210 on Saturday, how many does it typically make on a weekday,” a model prompted for direct reasoning will usually set up and solve the proportion correctly; one pushed straight to an answer is far more likely to guess a wrong number that merely looks plausible.
It tends to add less value on simple factual lookups or purely creative requests, where there’s no multi-step logic to walk through — asking a model to “think step by step” before naming the capital of a country doesn’t help, because there’s no chain to build.
How to use it in your own prompts
You rarely need to engineer anything elaborate. Adding a plain instruction — “work through this step by step before giving your final answer” or “explain your reasoning, then conclude with a one-line answer” — is usually enough to trigger the behavior. For tasks you run repeatedly, like analyzing data or checking a document for a specific kind of error, you can go further and specify the exact steps you want checked in order, which both improves accuracy and makes the output easier to audit. The underlying lesson generalizes well beyond AI: a process that externalizes its intermediate steps, whether it’s a model or a person, catches more of its own mistakes than one that only reports a conclusion.
