Ask for the answer directly and the model gets one shot per token; ask it to "think step by step" and each written step becomes context for the next, letting it decompose problems, catch errors, and carry sub-results forward. This one prompt trick reliably boosts math, logic, and code performance.

Reasoning models bake this in — trained to produce long internal thinking before answering. A caveat worth knowing from the research: the written chain doesn't always reflect the model's actual computation, so read it as useful working notes, not a faithful log.