Chain-of-thought shifts a model from "answer directly" to "reason then answer." Classic form: add "Let's think step by step" to the prompt, or show a few-shot example that includes reasoning before the answer.
On reasoning-tuned models (Claude 4.6, GPT-5, o-series), explicit CoT instructions are often redundant -- the model does it internally. On smaller or non-reasoning models, CoT still matters and can double accuracy on math and logic.
Example Prompt
Problem: A shop sells apples at 3 for $2 and oranges at 5 for $3. What does it cost to buy 12 apples and 15 oranges?
Think step by step, then give the final price.When to use it
- Multi-step math, logic, or planning problems
- The task requires combining multiple pieces of information
- You're working with a smaller or non-reasoning model
- You need the reasoning trace for debugging or auditing
When NOT to use it
- Simple lookup or classification -- CoT just adds latency
- Using a reasoning model that already thinks internally (wastes tokens)
- Latency-sensitive endpoints where the reasoning isn't user-facing value