What Is Chain-of-Thought Prompting? A Simple Guide

Ever asked an AI a tricky question and got a confident answer that was simply wrong? One of the oldest fixes is to ask it to show its working, the way a teacher would with a maths problem. That's chain-of-thought prompting, and it's still worth knowing, though with newer tools it's a bit more nuanced than it used to be.

Chain-of-thought prompting in simple terms

Wikipedia describes chain-of-thought as a technique that lets a language model solve a problem as a series of intermediate steps before giving a final answer. Instead of jumping to a conclusion, the model writes out its reasoning first.

A Google research paper by Jason Wei and colleagues introduced the idea. They showed that prompting large language models with worked examples containing step-by-step reasoning improved performance on arithmetic, commonsense and symbolic reasoning tasks.

Chain-of-thought prompting: a before and after

Imagine you ask: "I sell cupcakes in boxes of six. A customer wants enough for twenty people, with two each. How many boxes?" A rushed answer might be off. If you add "work through it step by step", the model tends to lay out the total needed, divide by the box size, and round up. You can then see each step and spot a slip.

That visibility is half the benefit. When the reasoning is on the page, you can check it yourself.

Two ways to ask for step-by-step answers

• With examples. This is the original approach. You show a worked example with its reasoning, then ask your question. Wikipedia notes this makes it a few-shot technique.
• With a simple phrase. Wikipedia says a later finding showed that adding "Let's think step-by-step" also worked, so chain-of-thought could be used with no examples at all.

If the example-based method interests you, look for my separate explainer on few-shot prompting, which covers how to write good samples.

How to use chain-of-thought prompting yourself

1. Say what you want first. State the problem clearly, including any numbers or rules.
2. Ask for the steps. For example, "Think through this step by step before giving your final answer."
3. Ask for a clear final line. Tell it to finish with a short answer you can copy, so you don't have to dig through the reasoning.
4. Check the steps. A wrong step can still lead to a confident answer, so read the logic, not just the result.
5. Try it twice for important work. Wikipedia describes self-consistency, where several attempts are run and the answer reached most often is chosen. You can copy the spirit by asking the same question a couple of times and comparing.

For more everyday phrasing ideas, see my collection of ChatGPT prompt examples.

Chain-of-thought prompts you can adapt

Here are a few phrasings that work in most chat assistants, written as plain sentences you can adjust. Ask it to list the facts it has before it calculates anything. Ask it to name any assumptions it's making. Ask it to check its own answer against the original question before it finishes. Each one nudges the model to slow down and show something you can inspect.

For longer jobs, Google's documentation suggests another route: break a complex task into several prompts, so each step's output feeds the next. That gives you a natural checkpoint after every stage, which is easier to review than one long block of reasoning.

When chain-of-thought prompting helps most

Wikipedia cites a meta-analysis of many studies that found large gains primarily on mathematical, logical and symbolic reasoning problems, with much smaller gains elsewhere. In business terms, think of tasks like:

• Working out quotes, margins or schedules with several moving parts.
• Comparing options against a list of rules.
• Planning a project in the right order.
• Checking whether a piece of logic in a spreadsheet formula makes sense.

When chain-of-thought prompting doesn't help

It isn't a cure-all. Wikipedia notes that added reasoning steps increase response times and that chain-of-thought is not always better than a simple prompt for factual or low-complexity questions. It also reports that on some tasks, where deliberation hurts even human performance, it reduced the accuracy of some advanced models.

So for a quick rewrite of a sentence or a plain factual question, skip it. You'll just get a longer answer.

Do newer AI tools still need chain-of-thought prompts?

Often less than before. Google's Gemini documentation says its models automatically generate internal thinking text to improve reasoning, so it's generally not necessary to have the model outline its reasoning steps in the response. It adds that a simple request such as asking the model to think hard before answering can help on heavy tasks, with some extra thinking time.

Anthropic's guidance says similar things about its own thinking features: the model can decide dynamically when and how much to think. If you're using a standard chat assistant, there's no harm in asking for steps on a tricky problem. If you're using a model built to reason, you may not need to.

Whichever tool you use, remember that a neat explanation isn't proof. Reasoning written by an AI can look convincing and still contain an error, which is one reason I always check the numbers myself. If you'd like help setting up sensible AI habits across your business, take a look at my work as an AI consultant.

Frequently asked questions

What is chain-of-thought prompting?

It's a technique where you get an AI model to work through a problem in intermediate steps before giving its final answer.

Who came up with chain-of-thought prompting?

A Google research paper by Jason Wei and colleagues introduced the term, showing step-by-step examples improved reasoning tasks.

What phrase triggers chain-of-thought?

Wikipedia says adding the phrase "Let's think step-by-step" was found to work without any examples.

Does chain-of-thought always improve answers?

No. It helps most on maths, logic and symbolic problems, and can slow things down or even hurt accuracy on simple or intuitive tasks.

Do I need it with newer AI models?

Often less. Google says its Gemini models generate internal thinking automatically, so spelling out steps is generally not necessary.

Can I trust the steps the AI shows?

Not blindly. The steps can look sensible and still contain a mistake, so check the logic and the numbers.