What is chain of thought prompting?
Chain of thought prompting (CoT prompting) means asking a language model to write out intermediate steps before it states an answer. Paste this into Claude, ChatGPT, Gemini or Whizi and fill the brackets.

Solve this problem: [your problem]. Work through it step by step and show each calculation. Then give the final answer on its own line that starts with "Answer:". If a step depends on an assumption, state the assumption.
The name comes from a 2022 paper, covered below. A "chain of thought" is the written run of steps between the question and the answer. Because a model produces text one piece at a time, each step it writes becomes input for the next one. Skip the steps and it must jump to the answer in one move, which is where multi step problems go wrong.
What did the 2022 Wei et al. paper show?
Wei and eight co authors at Google showed that a few worked examples with reasoning steps made large models much better at multi step tasks. The paper, "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models" (Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le and Denny Zhou), went on arXiv on 28 January 2022.
The prompt held a handful of example questions, each answered with its steps written out, followed by the real question. Experiments on three large language models showed gains on arithmetic, commonsense and symbolic reasoning tasks. The abstract reports that a 540 billion parameter model given eight such examples reached the best published accuracy on GSM8K, a set of grade school maths word problems, ahead of a finetuned GPT-3 with a verifier.
Two limits belong in any honest summary. The abstract describes the ability as one that emerges in large enough models, so small models gained little at the time. And the examples were written by hand for each task, which is why people looked for a way to drop them.
Does model size matter for chain of thought?
Yes, in the original work. Google Research's write up of the paper (11 May 2022) says models of sufficient scale, about 100 billion parameters, can solve complex reasoning problems this way, and that standard prompting gave relatively flat scaling curves on MultiArith and GSM8K. It reports a new best score of 58% on GSM8K for the 540 billion parameter PaLM with chain of thought. Smaller models tend to write chains that read well and lead nowhere.
Current chat models are far past that size, so the threshold rarely bites in practice. It still explains why a tiny on device model can fail with the same prompt that works on a large one.
Does "think step by step" work without examples?
Yes, on the models tested in 2022 it did, by a wide margin. Kojima, Gu, Reid, Matsuo and Iwasawa posted "Large Language Models are Zero-Shot Reasoners" on 24 May 2022. They put the sentence "Let's think step by step" before each answer, with no worked examples at all.
Those figures are from the paper's abstract and used InstructGPT (text-davinci-002); the authors report similar gains on the 540 billion parameter PaLM. Both numbers are old now. Treat them as proof that the step line mattered then, not as a gain you'll get today.
Zero shot CoT is the form most people mean by "CoT prompting" in everyday use: one added instruction, no examples. The few shot form from Wei et al. still wins when you need the steps in a fixed shape, such as a table or a checklist.
What is the difference between chain of thought and prompt chaining?
Chain of thought is one prompt that asks for visible steps. Prompt chaining is several separate prompts, where the output of one becomes the input of the next. TechTarget's explainer draws the same line.
Microsoft's .NET documentation (updated March 2026) shows two ways to write CoT: an instruction that tells the model to break the task into steps and output each result, or an example that lays out the steps as a template with blank headings for the model to fill. The second fixes the layout, which suits reports and comparisons.
Use CoT when one model can reason through the task in a single reply. Use a chain of prompts when the stages need different instructions, such as research, then outline, then draft. The two combine well: each prompt in a chain can ask for its own steps.
Chain of thought examples: three real cases, with and without
These three cases are written for this page, not taken from a model test. Each shows the plain prompt, the chain of thought prompt and what a good answer must contain.
Case 1, a discount sum. Plain prompt: "A cafe sells muffins at $3.50 and coffees at $2.75. A group buys 4 muffins and 5 coffees and gets 10% off the total. What do they pay?" Chain prompt: add "Work out each line total, then the subtotal, then the discount, then the final price." A good answer has four steps: $14.00 for muffins, $13.75 for coffees, a $27.75 subtotal, and $24.98 after the discount, rounded from $24.975. Without the steps, the usual failure is discounting only one item.
Case 2, an ordering puzzle. Plain prompt: "Ann is taller than Bo. Cy is shorter than Bo. Di is taller than Ann. Who is shortest?" Chain prompt: add "Write each comparison as a line, then build the full order from tallest to shortest." A good answer lists the order as Di, Ann, Bo, Cy, so Cy is shortest. The written order is the part you can check.
Case 3, a percentage trap. Plain prompt: "A jacket goes from $80 to $60, then rises 25%. Is it back to $80?" Chain prompt: add "Calculate the price after each change before you answer." A good answer finds $60 times 1.25 is $75, so the jacket is $5 short. The trap is that the 25% rise is measured on $60, not on $80.
In all three, the check is the same: the final line has to follow from the lines above it. If a model gives the right answer with steps that don't lead to it, it's luck. Don't count it.
Do reasoning models still need chain of thought prompting?
Mostly no. A reasoning model writes hidden steps before its reply, so asking for steps again adds cost and rarely adds accuracy. OpenAI's reasoning best practices have a section headed "Avoid chain-of-thought prompts", and it says prompting these models to "think step by step" or "explain your reasoning" is unnecessary. It adds that the instruction may not help and can sometimes make results worse.
Independent testing points the same way. Meincke, Mollick, Mollick and Shapiro (Wharton, 8 June 2025) found CoT "often results in only marginal, if any, gains" for reasoning models, and that it significantly raises the time and tokens a reply takes. For models that do not reason by default, they found a small average gain, with answers that vary more from run to run.
Anthropic's prompting guide frames manual CoT as a fallback for when thinking is off. It also says a general instruction such as "think thoroughly" often beats a hand written step by step plan. Read what is a reasoning model for how the hidden steps work and what they cost.
The practical rule as of November 2026: with a reasoning model or a thinking setting on, ask for the result and set the effort level. With a fast model, or any time you want the steps visible for checking, ask for CoT.
What do the sources say, and when?
Five primary sources cover the whole arc, from the first paper to current vendor advice. The table lists what each one states, and nothing it does not.
| Source | Date | What it states | What it means for your prompt |
|---|---|---|---|
| Wei et al., arXiv 2201.11903 | 28 Jan 2022 | Worked examples with steps lift large models on arithmetic, commonsense and symbolic tasks | Paste one solved example when format matters |
| Kojima et al., arXiv 2205.11916 | 24 May 2022 | "Let's think step by step" took MultiArith from 17.7% to 78.7% | One added line is enough on older models |
| Turpin et al., arXiv 2305.04388 | 7 May 2023 | Stated reasons can hide the real cause; accuracy fell by up to 36% when prompts were biased | Check the steps, do not trust them |
| Meincke et al., arXiv 2506.07142 | 8 Jun 2025 | Marginal gains on reasoning models, more time and tokens | Drop the line on thinking models |
| OpenAI reasoning best practices | Read Nov 2026 | Step by step prompts are unnecessary for reasoning models | Ask for the result |
| Anthropic prompting guide | Read Nov 2026 | Manual CoT is a fallback when thinking is off | Use it only with thinking off |
The table has a gap on purpose. No source above gives a CoT accuracy gain for any current model, so any number you read for today's Claude, GPT or Gemini needs its own test. The papers also used older models, and none of them tested a Whizi plan.
12 chain of thought prompts you can copy
These twelve prompts cover the forms people search for: zero shot, few shot, structured, self checking and the cases where you want the steps for review. Replace each [bracket] with your detail.
- Zero shot: "[your question] Let's think step by step."
- Zero shot with a clean finish: "[your question] Think through it step by step. End with one line that starts with Answer:."
- Few shot: "Example question: [solved problem]. Example steps: [its steps]. Example answer: [its answer]. Now solve this one the same way: [new problem]."
- Tagged answer: "Reason through [problem] first. Then put only the final answer inside <answer> tags so I can copy it."
- Least to most: "Break [big problem] into the smallest sub questions, answer them in order, and use each answer in the next one. Then answer the original question."
- Assumptions first: "Before you solve [problem], list every assumption you need. Then solve it step by step and mark where each assumption is used."
- Two routes: "Solve [problem] two different ways. If the two answers differ, find the step that is wrong and say which answer you trust."
- Self check: "Solve [problem] step by step. Then re check each step and the units. Tell me if you changed anything."
- Spot the error: "Here is my working: [paste]. Go through it line by line and stop at the first line that is wrong. Explain why in one sentence."
- Plan then act: "I need to [goal]. Write a numbered plan with the reason for each step. Wait for me to say go before you write anything else."
- Decision with criteria: "Choose between [option A] and [option B] for [my situation]. List my criteria, score each option on each one with a sentence of reasoning, and then recommend one."
- Teach it back: "Explain how you got to [answer] as if to a student who got a different answer. Name the step where a person is most likely to slip."
Prompts 1 to 4 work in any chatbot. Prompts 7, 8 and 12 also work as a second pass on an answer you already have. For more starting points across everyday tasks, see best ChatGPT prompts and Claude prompts, and for the basics of a good brief, prompt engineering for beginners.
What mistakes ruin chain of thought prompts?
Four mistakes cause most bad results, and each has a short fix.
- Asking for the answer first and the steps after. The steps then explain a guess. The fix: say "steps first, then the answer".
- Using it on a task with no steps, such as a lookup or a rewrite. The fix: skip it, since it only adds length.
- Stacking it on a reasoning model. The fix: remove the step line and set the model's effort instead.
- Trusting a tidy chain. The fix: read the steps and check each number against a source or a second calculation.
Is the model's chain of thought really how it got the answer?
Not always. Turpin, Michael, Perez and Bowman (NeurIPS 2023) showed that written reasoning can leave out what actually drove an answer. When they reordered multiple choice options so the right answer was always "(A)", models shifted toward "(A)" while the explanations never said so. Accuracy fell by as much as 36% across 13 BIG-Bench Hard tasks for GPT-3.5 and Claude 1.0, and the models often built a reasonable sounding case for the wrong choice.
Those were 2023 models, so today's will differ. The lesson stands: a chain of thought is evidence you can inspect, not a record of what happened inside. Why AI gets things wrong and what is an AI hallucination explain the wider problem, and how to fact check AI answers gives a two minute routine.
Five questions to ask an AI about chain of thought
These prompts use a model to test your own understanding of the topic.
- "Explain chain of thought prompting to me using a problem from [my field]. Show the answer with and without steps."
- "Give me three problems of increasing difficulty in [subject] where step by step reasoning should help, and one where it should not. Say why for each."
- "Here is a chain of thought a model wrote: [paste]. Does each step follow from the last? Mark any step that adds a fact without a source."
- "Rewrite my prompt [paste] as a zero shot CoT prompt, a few shot prompt and a plain prompt. Tell me which I should try first for [task]."
- "Which parts of this task would a reasoning model do without being told to reason? [describe task]"
Where does Whizi fit for chain of thought prompting?
Whizi is one app instead of a pile of single purpose tools: Pro carries Claude, GPT and Gemini together, and its first week is $1.99. That makes the test above easy to run. Send prompt 1 to one model and a plain version to another, or use side by side Compare on Pro to put one prompt to two models and read both.
Web search is on by default in each chat, which helps when a step depends on a current fact. The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card. A dedicated research notebook or a vendor's own API playground gives finer control of reasoning settings than a chat app does.
- Use chain of thought for tasks with steps, not for lookups or rewrites
- Ask for the steps before the answer, and pin the answer to a final line
- Turn the step line off when a reasoning model or thinking setting is on
- Read the steps and check each number, since a tidy chain can still be wrong
- Test any claimed gain on your own task before you rely on it
Frequently asked questions
What is chain of thought prompting in simple words?
It means telling the model to show its steps before the answer. A model that writes "first this, then this" makes fewer slips on maths, logic and planning than one that jumps straight to a result. The 2022 Wei et al. paper introduced it with worked examples, and Kojima et al. showed that "Let's think step by step" works with none.
Is CoT prompting still useful in 2026?
Yes with fast models and when you want visible steps to check, and rarely with reasoning models. OpenAI's guide says step by step prompts are unnecessary for reasoning models, as of November 2026, and a 2025 Wharton study found marginal gains for them. Test both versions on your own task.
Does chain of thought cost more tokens?
Yes, because the steps are extra text the model writes and you are billed for. The Wharton study found CoT significantly raised time and tokens on reasoning models. Paste a reply into the token counter to see its size and an estimated API price before you scale a prompt.