Chain of Thought

What is chain of thought?

You've probably seen it: ask an AI a hard question and it blurts out an answer that's wrong. Chain of thought (CoT) tells it not to rush. Like a student showing their work in math class, the model lays out each reasoning step before reaching a conclusion. Oddly enough, just adding "think step by step" is enough to lift accuracy by a lot.

Why does it work?

It turns skipping into solving
A model generates one token at a time. Skip the middle steps and it's like guessing — easy to get wrong. CoT forces it to spread the process out, so mistakes get caught in the middle.
The steps are the scratch paper
Each step written into context becomes input for the next one. The clearer the scratch work, the more reliable the final answer.

How do you use it?

Simplest: add one line
"Reason step by step" or "think first, then answer" — costs nothing to try.
Next level: show an example
Put a worked example with its reasoning in the prompt, and the model will follow that rhythm.
Harder: let it draft a few times
Have the AI think through several versions internally and pick the best (the "self-consistency" trick) for steadier results.

What did it change?

Chain of thought is one of prompt engineering's classic tricks, and it's also the idea behind today's "reasoning models" — models that think internally before answering. It turns AI from fast-talking into think-before-you-speak.

Bottom line: chain of thought makes AI turn mental math into shown work, walking you through each step so the answer comes out more reliable.