Ask the Model to Show Its Work (Chain of Thought)
What it is
Chain-of-thought prompting means asking the model to lay out its intermediate steps before giving a final answer, instead of jumping straight to the result. For a problem with several moving parts — a calculation, a decision with trade-offs, a multi-step plan — you ask it to reason through the steps first, then state the answer.
The term comes from a 2022 research paper, "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models," by Jason Wei and colleagues. The technique is a prompting move, not a setting: you're changing what you ask for — the steps, then the answer — not flipping a switch inside the model.
Time & frequency
One extra line in your prompt. Use it whenever the question has more than one step, or when you need to check the reasoning.
How to do it
- 1Write your question as usual — the actual problem you want solved.
- 2Add an instruction to reason first: "Work through this step by step, then give your answer."
- 3Ask it to separate the two parts so you can read them apart: "Put your reasoning in <thinking> tags and your final answer in <answer> tags."
- 4For a worked problem, show one example that includes the steps, so Claude copies the pattern of reasoning, not just the format.
- 5Send it to Claude at claude.ai and read the steps, not just the answer. If a step is wrong, you've found exactly where it went off.
- 6If the reasoning is sound but long, ask for a shorter chain next time: "Reason briefly, in three or four steps."
Why it works
Asking for intermediate steps changes what the model produces: instead of one leap to an answer, it lays out the reasoning that leads there. That gives you something to check — you can see each step and catch the one that's off — and it keeps a multi-step problem from being answered in a single guess.
The technique was named and described by Wei and colleagues in 2022 as "chain-of-thought prompting": prompting the model to produce a series of intermediate reasoning steps. This write-up teaches the prompting move — ask for the steps, then the answer — not any benchmark score. Anthropic's own guidance describes the same manual move: "you can still encourage step-by-step reasoning by asking Claude to think through the problem," using tags like <thinking> and <answer> to separate reasoning from the final output.
Sources
- Wei, J., et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arXiv:2201.11903 — names and describes chain-of-thought prompting.
- Anthropic — Prompting best practices (Claude docs): manual chain-of-thought prompting — asking Claude to think through a problem step by step, separating <thinking> from <answer>.
Common mistakes
- Using it for everything. A one-step lookup doesn't need a chain — you'll just wait longer for the same answer.
- Not separating the steps from the answer. If reasoning and answer blur together, you can't check the work. Ask for tags.
- Reading only the final line. The whole point is to scan the steps and spot where it went wrong.
- Trusting a neat-looking chain. Plausible steps can still be wrong — check the one the answer hinges on.
Practice it now
Try a chain-of-thought prompt
Ask Claude to show its steps before the answer, then check them. Free, no signup.
- Write a question with more than one step.
- Add: "Work through it step by step, then answer."
- Ask for reasoning in <thinking> tags and the answer in <answer> tags.
- Send it to Claude at claude.ai and read the steps, not just the answer.
Related skills
Coherent Education
You just got the skill for free. The coached version is where it sticks.
seven live tracks, 58 coach-guided lessons, and games that train the skill — with a rep log so you can watch the work add up.
See the tracks