Chain-of-thought prompting is a technique where you ask an AI model to reason through a problem step by step before giving its final answer, which measurably improves accuracy on tasks that need multiple steps. Instead of jumping straight to a conclusion, the model writes out its intermediate reasoning, and that visible working tends to catch errors it would otherwise make. The classic trigger is adding a phrase like "let us think step by step" to your prompt. This guide explains why it works, when to use it, and how it relates to the reasoning models common in 2026.
Why reasoning out loud helps
A language model generates one piece of text at a time, each conditioned on what came before. If you force it to lay out the steps, each step becomes context that guides the next, so the model is far less likely to skip a hop or make a careless arithmetic slip. The reasoning acts as scratch paper.
For a one-fact question, this does nothing useful. For a word problem, a logic puzzle, or a multi-part analysis, it can turn a wrong answer into a right one. The harder and more sequential the task, the bigger the gain.
How to use it
| Prompt style |
What you do |
Best for |
| Zero-shot CoT |
Add let us think step by step |
Quick boost, any task |
| Few-shot CoT |
Show worked examples with reasoning |
Consistent format, tricky tasks |
| Ask to verify |
Have it check its own steps |
Catching arithmetic or logic slips |
| Structured CoT |
Request numbered steps then answer |
Auditable, parseable output |
The simplest version costs nothing: just ask for the steps. For repeatable quality, show a couple of worked examples so the model copies your reasoning style. This sits inside the broader craft of writing prompts that work.
Chain-of-thought versus reasoning models
By 2026, many models do this kind of reasoning internally without being told. These reasoning models think before answering by default, which means you prompt for chain-of-thought explicitly less often than you used to.
- With a reasoning model, you can often just ask the question and let it reason.
- With a standard model, the trigger phrase still helps on hard problems.
- Either way, asking to show the steps makes the answer auditable, which matters when correctness is important.
- For sensitive work, the visible reasoning lets you spot where it went wrong.
A close cousin is few-shot prompting, where examples guide the format; combining the two is common.
What to skip
- Do not use chain-of-thought for trivial lookups or simple rewrites. It adds tokens, cost, and latency for no benefit.
- Do not trust the displayed reasoning as a perfect account of how the model decided; it is a helpful artifact, not a guaranteed inner log.
- Do not paste enormous prompts when a reasoning model would handle the task natively. Match the technique to the difficulty.
The reasoning text can also be wrong while looking confident, so verify conclusions on anything that matters.
FAQ
What does chain-of-thought prompting do?
It makes a model spell out intermediate steps before answering, which improves accuracy on multi-step problems like math and logic by giving the model a structured path to the answer.
What is the magic phrase?
Something like let us think step by step. Adding it often triggers step-by-step reasoning. Showing worked examples (few-shot) makes the effect more reliable.
Do I still need it with reasoning models?
Less often. Many 2026 models reason internally by default, so you can ask the question directly. The technique still helps with standard models and when you want auditable steps.
Does it always improve answers?
No. It helps on hard, sequential tasks but wastes time and money on simple ones. The reasoning can also be flawed, so verify important results.
Where to go next
Learn how to write prompts that work, understand few-shot prompting, and see what a system prompt is.