Chain of thought prompting is a prompting technique in which you ask an AI model to reason step by step before it gives an answer. Instead of presenting a conclusion straight away, the model writes out its intermediate steps. For complex questions, that often produces better results, and results you can check.
The technique was introduced in 2022 by Google researchers led by Jason Wei. They showed that language models are considerably more accurate on arithmetic, logic and reasoning tasks when they spell out their intermediate steps explicitly. Since then, chain of thought has become one of the most widely used techniques in prompting.
How does chain of thought prompting work?
Chain of thought prompting works by getting the language model to generate intermediate steps instead of jumping straight to a final answer. With a standard prompt, you often get only a conclusion. With a chain of thought prompt, you also see how the model arrives at that conclusion.
- Facts: are the starting points correct?
- Logic: do the steps follow?
- Conclusion: does it follow from the steps?
The written-out steps make the output easier to check, not automatically correct.
The difference is like that between someone who says “the answer is 42” and someone who shows you the calculation behind it. The written-out reasoning lets you, as the user, judge whether the logic holds and see exactly where it might go wrong. A model predicts patterns and does not check facts. Chain of thought only makes that prediction more visible.
There are two common variants of chain of thought prompting. Which one you choose depends on how complex your task is and how much time you have to prepare your prompt.
Zero-shot chain of thought prompting
Zero-shot chain of thought prompting is the simplest variant. You add a short instruction to your prompt, such as “Think about this step by step” or “Explain how you arrive at your answer”. The model gets no examples, only the instruction to show its reasoning.
This variant is quick and practical. You do not need to prepare examples and can apply it to any question straight away. On simple reasoning tasks, the results are already noticeably better than without a chain of thought instruction.
Here is an example. Say you ask ChatGPT: “A project team has 12 tasks. Three team members split the tasks equally. Then 6 more tasks come in, which are split between two team members. How many tasks does each team member have now?” Without a chain of thought instruction, you sometimes get a wrong answer. Add “Work this out step by step” and the model writes out each step and almost always reaches the right answer.
Few-shot chain of thought prompting
Few-shot chain of thought prompting goes a step further. You give the model one or more examples in which you show what a line of reasoning looks like. The model then follows that pattern for the new question.
This variant is particularly effective for tasks where the reasoning style you want is not obvious, such as situations in which you want to follow a specific decision framework, or in which the reasoning has to weigh several perspectives. By showing a specific example, you steer the model in the right direction.
The downside is that few-shot prompts take more preparation. You have to write a good example yourself, and the quality of that example directly affects the result. In practice, few-shot chain of thought prompting is most valuable for recurring tasks, where the effort of writing a good example pays for itself.
Chain of thought prompting in practice
Most articles explain chain of thought prompting with maths problems. That makes sense, because the effect is easiest to measure on arithmetic tasks. But the technique is at least as valuable for professional tasks that call for weighing options or analysis.
Example: an analysis
Chain of thought prompting in an analysis helps you check the model’s reasoning. Say you ask an AI model to compare three suppliers on price, delivery time and quality. Without a chain of thought instruction, you get a conclusion (“Supplier B is the best choice”) with no justification.
Add “Weigh each factor separately and explain how you arrive at your recommendation” and you see exactly which assumptions the model makes. Perhaps it weighs price more heavily than quality, while for your situation it should be the other way round. That visibility is the difference between trusting blindly and judging deliberately.
Example: a business decision
Chain of thought prompting for a business decision forces the model to work out arguments for and against. Say you ask Claude whether it is wise to roll out a new software package for your team. A standard answer lists advantages and ends with “it depends on your situation”.
With a chain of thought instruction such as “Go through the pros and cons step by step, weigh them against each other and reach a concrete recommendation”, you get a structured analysis. The model covers costs, implementation time, learning curve and expected productivity gains separately. That makes the output usable as a starting point for a real decision, rather than a noncommittal overview.
When to use chain of thought prompting
Chain of thought prompting does not work equally well for every type of task. The technique makes a noticeable difference for some assignments and adds little or nothing to others. Telling the two apart saves time and keeps outputs from becoming needlessly long.
Tasks where the technique helps most
Chain of thought prompting helps most with tasks that require several steps or trade-offs: arithmetic problems, logic puzzles, comparisons between options, risk analyses and policy decisions. What they have in common is that the answer cannot be reached in a single step.
The more complex the task, the bigger the difference from standard prompting. As soon as several variables are involved or a trade-off is needed, you will notice the quality of the output go up. Professionals who ask AI good questions often combine chain of thought prompting with other techniques, for example with a role description (“You are a financial analyst. Go through the following data step by step…”) or with a specific output format.
Tasks where the technique adds nothing
Chain of thought prompting adds nothing to simple tasks. For factual questions, short writing tasks or creative assignments, the technique does not produce better results; the model just makes the answer needlessly long without improving its quality.
For example, if you ask “Write a short email moving a meeting to Thursday”, step-by-step reasoning is unnecessary. The model does not need to work through intermediate steps; it only has to generate a text.
The rule of thumb is this: if you could phrase the answer yourself in one step, chain of thought prompting serves little purpose. As soon as you notice that you would need to think about the right approach yourself, it is a good moment to give the model the same room to reason. Longer answers are not necessarily better answers either. A model that reasons at length but is on the wrong track produces more text without more value.
Chain of thought versus prompt chaining
Chain of thought prompting and prompt chaining are often confused, but they are very different techniques. The difference lies in where the reasoning takes place: within one prompt or spread across several prompts.
With chain of thought prompting, everything stays within one prompt. You ask the model to reason step by step and reach a conclusion in a single answer. The whole reasoning process sits in the same response. With prompt chaining, you split the task into several separate prompts. The output of prompt 1 becomes the input of prompt 2, and so on. Between steps, you can adjust, validate or correct the output.
The choice between the two depends on the complexity and risk of the task. For well-defined reasoning tasks where you want to see the whole process at once, chain of thought works fine. For longer production workflows, where you need to check the intermediate results at each step or where each step has its own output format, prompt chaining is stronger. The techniques do not exclude each other: you can easily apply chain of thought within one step of a prompt chain, for example in the step that involves the heaviest trade-off.
Chain of thought prompting with reasoning models
Chain of thought prompting was originally something you had to trigger yourself through your prompt. That is changing. Recent reasoning models have chain of thought reasoning built into their architecture by default.
OpenAI’s o-series (o1, o3, o4-mini) was long the best-known example; o3 and o4-mini have since been removed from ChatGPT. Reasoning models are trained to work through an internal reasoning process before they give an answer. You see this in the interface as a “thinking” phase in which the model builds up its own chain of logic. As a user, you no longer need to give an explicit chain of thought instruction; the model does it automatically. When it launched, DeepSeek R1 went a step further and showed the full reasoning process in the output, including moments where it corrected itself. As of 13 October 2026, DeepSeek’s current models are DeepSeek-V4.1-Flash (since 10 September 2026) and V4-Pro, and in the web chat you switch on reasoning with DeepThink.
Chain of thought reasoning is also built into the current Claude models. In its model documentation, Anthropic lists adaptive thinking for Claude Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1, and for Opus 5.5 and Fable 5.1 that reasoning step is always on. The same applies to these models as to the other reasoning models in this section: the reasoning layer is already there, and the gain lies in a clear instruction and in checking the outcome.
Recent research from the Wharton Generative AI Labs (Meincke, Mollick et al., 2025) shows that chain of thought instructions add hardly any value with reasoning models and in some cases even give slightly worse results, while costing more compute time. For non-reasoning models, the technique still delivers a modest improvement on average. In my own work with Claude Opus 4.7, I noticed that an explicit chain of thought instruction sometimes confused the model more than it helped; the output got longer and wordier without the substance getting any sharper. With models like these, short, clear instructions work better than lengthy explanations up front.
For professionals, this is a relevant shift. With reasoning models, the skill moves from “writing the right chain of thought prompt” to “critically assessing the reasoning shown”. You no longer need to tell the model how to reason, but you do need to judge whether the steps it shows hold up. For models without a built-in reasoning layer, chain of thought prompting remains a valuable technique that you trigger yourself.
Limitations of chain of thought prompting
Chain of thought prompting improves output on complex tasks, but it is no guarantee of correct answers. There are two limitations you need to watch out for as a professional.
Convincing but incorrect reasoning
Chain of thought prompting can lead to reasoning that is convincing but wrong. The model generates text that looks logical, step by step, but is built on incorrect assumptions. A language model recognises patterns in text and predicts the most likely next word; it does not check its own logic the way a person does.
In practice, this means you always have to go through the reasoning, especially for tasks where mistakes have consequences. A financial overview that looks neat but rests on a wrong assumption is more dangerous than a visibly messy answer. The written-out steps let you find mistakes, but only if you actually read and assess them.
Longer answers are not always better answers
Chain of thought prompting produces longer answers. Sometimes that is exactly what you need, but it can also add noise. A model that writes out seven steps where three would have done makes it harder to find the core.
On simple tasks, chain of thought prompting can even make the output worse. That is not an anecdotal observation but a finding from the original Wei et al. study (2022): for questions that require only one step, chain of thought brought no improvement and sometimes a slight decline. So use the technique selectively, not by default.
Chain of thought prompting within a checkable workflow
Chain of thought prompting fits well with a way of working in which you check AI output systematically before you use it. The written-out reasoning gives you specific points at which to assess whether the answer is right.
You can check the reasoning on three levels. Are the facts the model starts from correct? Are the intermediate steps logical? And does the conclusion follow from those steps? If one of the three does not hold, you know exactly where the problem lies and can make a targeted correction.
In my own workflow, I often use chain of thought as a reviewer step. Once I have a first answer, I send a second prompt asking the model to check its own reasoning step by step for assumptions and leaps in logic. That forces a more explicit justification than the first run did and catches mistakes before the output goes any further. The LearnLLM ChatGPT Course sets out this kind of workflow step by step, including fixed review points and example prompts you can adopt in your own work.
Chain of thought prompting is therefore more than a trick for better answers. It is a way to make AI output checkable, which is exactly what professionals need when they deliver output under their own name. If you want to use this technique systematically in your work, LearnLLM’s ChatGPT Course gives you a complete way of working with repeatable workflows and review points.
Frequently asked questions about chain of thought prompting
Chain of thought prompting often raises further questions about how reliable the reasoning shown is and what the reasoning steps cost.
Is the reasoning a model shows its real reasoning?
Not always. The chain of thought a reasoning model shows does not necessarily reflect what the answer is actually based on. Anthropic studied this in April 2025 with Claude 3.7 Sonnet and DeepSeek R1 by giving the models hints that influenced their answers. The models often did not mention those hints in their reasoning. So read written-out reasoning as an aid for finding mistakes, not as proof that the answer is correct.
Do you pay for a model’s reasoning steps?
Yes, through the API you pay for a model’s reasoning steps, because they count as output tokens. Anthropic shows in each response how many of the billed output tokens were internal reasoning, and those thinking tokens also count towards the limit you set with max_tokens. A chain of thought instruction to a model without a built-in reasoning layer makes the answer longer and therefore more expensive.
What is self-consistency in chain of thought prompting?
Self-consistency is an extension of chain of thought prompting in which the model produces several lines of reasoning for the same question and you pick the answer that comes up most often. The technique comes from Google research in 2022, with Jason Wei as one of the authors. In practice, you ask the same question a few times in separate conversations and compare the outcomes. If the answers differ widely, that is a signal to check the reasoning more closely.
