Prompting & Everyday AI Use · Prompt Engineering

What Is Chain-of-Thought Prompting?

Chain-of-thought prompting is a technique where you ask an AI model to work through a problem step by step before giving a final answer, rather than jumping straight to a conclusion. This often improves accuracy on tasks involving logic, math, or multi-step reasoning.

Key takeaways

Chain-of-thought prompting asks the model to show intermediate reasoning steps instead of outputting only a final answer.
It tends to improve performance on problems that require multiple logical or arithmetic steps, such as word problems or multi-part analysis.
Simple phrases like 'think step by step' or 'explain your reasoning before answering' can trigger this behavior in many chat-based models.
Some newer models apply extended internal reasoning automatically, reducing the need to manually request step-by-step thinking.
Chain-of-thought output is not proof of correctness — the stated reasoning can still contain errors even when it looks logical.

Asking the Model to Show Its Work

Chain-of-thought prompting means instructing an AI model to break a problem into intermediate steps and reason through them before producing a final answer, rather than generating a conclusion immediately. Instead of asking “What’s 15% of 240 minus 12?” and taking whatever number comes back, a chain-of-thought version might add “work through this step by step” — prompting the model to first calculate 15% of 240, then subtract 12, showing each step along the way.

This matters because language models generate text sequentially, one piece at a time, using everything generated so far as additional context. When a model is pushed to lay out intermediate steps, those steps become part of the input it uses to generate the next piece of text — effectively giving it a scratchpad that can make the final answer more consistent and accurate, especially for problems with several dependent parts.

Why Step-by-Step Reasoning Improves Results

The core reason chain-of-thought prompting helps is that many tasks — word problems, multi-step logic puzzles, planning tasks with several constraints — are hard to get right in a single leap, even for people. Skipping straight to an answer means the model has to hold an entire chain of reasoning implicitly, without any of the “checkpoints” that make it possible to catch an error partway through. Writing out each step gives the model (and the person reading the output) a chance to verify each part individually.

This is why chain-of-thought prompting became a widely used technique for improving performance on reasoning benchmarks: problems that a model got wrong when asked for a direct answer were often answered correctly once it was prompted to reason step by step first. The effect is strongest on tasks with clear intermediate structure, like arithmetic, logical deduction, or breaking down a business problem into components.

More recent AI models have started incorporating extended reasoning as a built-in behavior — generating internal step-by-step analysis automatically for harder queries, sometimes labeled as an “extended thinking” or “reasoning” mode, without the user needing to explicitly request it. That said, explicitly asking for step-by-step reasoning still helps with many everyday models and use cases, particularly older or lighter-weight ones.

A Practical Example

Suppose you ask an AI to figure out how many weeks it will take to save $3,000 if you save $180 a month, accounting for a one-time $200 expense in month two. A direct-answer prompt risks the model conflating the steps and producing a wrong number. A chain-of-thought version — “walk through this calculation step by step, month by month” — encourages the model to track the running total explicitly, month by month, which is both easier to verify and less prone to compounding errors.

Bottom Line

Chain-of-thought prompting improves accuracy on reasoning-heavy tasks by having the AI reason through intermediate steps instead of jumping to a conclusion, and it’s a simple, free technique to try whenever a task involves math, logic, or multiple dependent parts — though it adds little value for simple lookups and doesn’t guarantee the reasoning shown is actually correct.

Build a Better Prompt

Turn a simple form into a clear, ready-to-copy prompt with our free Prompt Builder — no blank page required.

Go deeper

Important caveats

For simple factual or creative tasks, chain-of-thought prompting adds length without meaningfully improving quality.
A model's written reasoning doesn't always reflect the actual computation behind its answer, so it shouldn't be treated as a verified audit trail.

Frequently asked questions

Does chain-of-thought prompting work on every AI model?

It works best on capable, general-purpose language models handling reasoning-heavy tasks. Its effect is less noticeable on smaller or narrowly specialized models, and some newer models reason step by step internally without needing to be asked.

Should I always ask the AI to show its reasoning?

No. It's most useful for math, logic, planning, or multi-step analysis. For simple lookups or creative writing, it mainly adds unnecessary length to the response.

Is chain-of-thought the same as the model 'thinking'?

Not in a human sense. It's a text-generation pattern where the model produces intermediate reasoning tokens that statistically tend to lead to more accurate final answers, not evidence of conscious deliberation.

Sources

[1]OpenAI Help Center — OpenAI
[2]Anthropic Documentation — Anthropic

Written by Editorial Team

Last updated July 25, 2026

Get one well-sourced answer a week

No spam. Unsubscribe anytime.