Chain-of-thought prompting instructs the model to reason through a problem step by step before giving a final answer. A single instruction transforms a direct question into an explicit reasoning process.
"What is 15% of 240?"
The model generates an answer in one step. For multi-step math, it often makes arithmetic errors that are invisible in the output.
"Think through this step by step. What is 15% of 240?"
The model first writes: "15% = 15/100 = 0.15. 0.15 × 240 = 36." Then gives: 36. Each step feeds the next.
When a model generates an answer directly, it makes a single forward pass and produces the most statistically likely response for the whole answer at once. For multi-step problems, this often fails — the correct final answer requires correct intermediate steps, and the model has no working memory to verify them.
When you ask for step-by-step reasoning, the intermediate steps appear as tokens in the context. Each step is visible and informs the next. The model is doing the work out loud — and visible work is self-correcting in a way that hidden work is not.
Simple instruction — "Think step by step before answering." — the most basic form; works well for most analytical tasks with minimal prompt overhead
Explicit reasoning request — "Show your reasoning before giving the final answer." — useful when you need to inspect or verify the logic, not just the conclusion
Few-shot chain-of-thought — provide examples that show the full step-by-step reasoning process, not just the final answer; the model learns to follow that structure
Self-check instruction — "After reasoning through the problem, check your answer before responding." — adds a verification step that catches errors in the generated reasoning
Input: A train travels 60 km/h for 2 hours. How far does it go?
Reasoning: Distance = speed × time. Speed = 60 km/h. Time = 2 hours. Distance = 60 × 2 = 120 km.
Answer: 120 km.
Input: [Your question here]
Reasoning: [Model fills this in]
Answer: [Model fills this in]
The task requires more than one reasoning step.
The final answer depends on intermediate calculations or logical deductions.
The model produces consistently wrong or inconsistent answers on direct prompting.
You need to inspect or verify the reasoning, not just accept the answer.
The task is a simple factual lookup — no reasoning chain needed.
The task is format transformation: translation, rewriting, classification.
Latency is critical and the task does not benefit from step-by-step reasoning.
Chain-of-thought generates more tokens — it increases latency and cost. The intermediate steps are generated by the same model, so a flawed reasoning step leads to a flawed conclusion. The model can generate plausible-sounding but logically incorrect reasoning chains. Asking the model to think step by step makes errors visible, but does not prevent them. For high-stakes reasoning, always verify the intermediate steps, not just the final answer.
Ask the AI if you need help understanding chain-of-thought prompting, why step-by-step reasoning improves accuracy, which technique to use for a specific task, or what the limitations of chain-of-thought are in practice.