Chain-of-thought prompting: asking a model to reason step by step

What asking for step-by-step reasoning actually does

A prompt like "work through this step by step" or "show your reasoning before giving a final answer" asks the model to lay out intermediate steps rather than jump straight to a conclusion. For a multi-step problem, this can surface a mistake in an intermediate step that would otherwise be buried inside a single, confident-sounding final answer — and it gives you something concrete to check rather than only a result to trust.

When it helps, and when it is unnecessary

Step-by-step prompting tends to help most on problems with several dependent steps, several constraints to satisfy at once, or a calculation worth double-checking — a multi-part word problem, a decision with several factors, a piece of logic that has to hold together. A short, single-step question rarely benefits from it, and asking for it anyway just adds length without changing the answer. See choosing a model for complex reasoning, linked below, for how this connects to picking a model built for exactly this kind of task.

What it does not guarantee

Asking a model to reason step by step does not make the answer correct, and a confident, well-structured chain of steps can still reach the wrong conclusion — the extra steps make an error easier to spot, not impossible to make. This is consistent with why AI hallucinates, linked below: a model can produce fluent, plausible-looking reasoning that is still wrong, so a step-by-step answer is worth checking on anything that matters, not taken on faith because it looks methodical.

Questions people ask

Does chain-of-thought prompting make the answer correct?

No — it does not make the reasoning correct, and a step-by-step answer can still reach a wrong conclusion. It tends to make a mistake easier to spot in the intermediate steps, which is different from preventing the mistake.

When should I ask a model to show its reasoning?

On problems with several dependent steps or constraints, where an intermediate mistake would otherwise be hidden inside a single final answer. A short, single-step question rarely needs it.

Is this the same as using a reasoning-focused model?

Related but not identical — this guide is about how you phrase a prompt to any model; choosing a model for complex reasoning, linked below, is about which model is built to work through steps by default. The two can be combined.