Chain-of-thought prompting asks a model to show its reasoning before giving an answer. On problems needing several steps, that alone improves accuracy substantially.
The reason is mechanical. A model produces one token at a time with fixed computation per token, so demanding an immediate answer gives it no room to work. Reasoning aloud creates that room.
Think of it like this. Think of mental arithmetic against working on paper. The sums are identical. Being allowed to write the intermediate steps is what stops you losing track halfway.
How it works
- Ask for the reasoning first, then the answer.
- Each step conditions the next. Generated reasoning becomes part of the context the answer is produced from.
- Zero-shot works. Adding "think step by step" is often enough.
- Few-shot works better. Showing worked examples teaches the reasoning style you want.
When to use it
- Multi-step arithmetic and logic, where the gains are largest.
- Anything with intermediate conclusions, such as debugging or planning.
- Not for simple lookups. It adds tokens, latency and cost for no benefit.
- Weigh the cost. Reasoning tokens are billed like any other, and on a high-volume endpoint that adds up.
Common mistakes
- "The reasoning shows how the model decided." It is generated text that leads to an answer, not a transcript of an internal process. It can be wrong while the answer is right, and the reverse.
- "It always helps." On simple tasks it adds cost and can talk a model out of a correct first instinct.
- "It works equally on any model." The benefit grows with model size. Small models often produce reasoning that does not help.
