Chain of thought: definition, examples of prompts and difference with prompt chaining (2026)

Article published on September 16, 2024, rewritten on October 24, 2026 and completed on November 14, 2026, established from the original research publications (Wei et al. 2022, Kojima et al. 2022, Wang et al. 2022, Dhuliawala et al. 2023) and the prompt guides of OpenAI and Anthropic consulted on that day.

The chain of thought is a prompting technique that requires an AI to reason through intermediate steps before providing its final answer. Instead of jumping directly to the result, the model presents the data, performs calculations or deductions one by one, and then concludes. In logic, computation, or decision-making problems, this significantly reduces errors and makes the reasoning readable. By 2026, the technique had changed its status: reasoning models now performed this task automatically. Here's what a chain of thought is, how to write it, when it's still useful, and how it differs from few-shot, prompt chaining, and chain of verification.

The essentials in five lines

  • Definition: have the model produce the steps of its reasoning before the answer, in a single exchange.
  • Two ways to trigger it: an instruction of the type "reason step by step" (zero-shot) or solved examples with their reasoning (few-shot).
  • Origin: popularized in 2022 by a team from Google (Wei et al.), then by the formula "Let's think step by step" (Kojima et al.).
  • In 2026: With reasoning models (ChatGPT's "reflection" mode, Claude's extended reflection, Gemini Thinking), the "step-by-step" instruction becomes unnecessary; the prompt's work shifts towards context, success criteria, and verification.
  • Not to be confused: Prompt chaining links several distinct prompts together, while chain of verification checks a response with control questions.

What is a chain of thought?

Chain of Thought (CoT) is a prompt engineering technique. It starts from a simple observation: a language model produces its answer word by word, and each written word serves as a starting point for the next. When asked to write the intermediate steps first, it has these steps to build the conclusion. When asked for the answer immediately, it must "guess" the result all at once, which often fails as soon as the problem involves several operations.

The term comes from the article "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models," published in January 2022 by Jason Wei and his colleagues at Google. They showed that by adding a few solved examples along with their reasoning to the prompt, a large language model made significant progress in arithmetic, symbolic reasoning, and common sense. With eight such examples, their 540-billion-parameter PaLM model achieved the highest score at the time on GSM8K, a benchmark for school-level mathematical problems.

A few months later, Takeshi Kojima and his co-authors demonstrated that providing examples wasn't even necessary. Simply adding the phrase "Let's think step by step" increased the accuracy of the OpenAI model they were testing from 17,7% to 78,7% on the MultiArith benchmark, and from 10,4% to 40,7% on GSM8K. This marked the birth of the "zero-shot" chain of thought, the most well-known version.

The two ways to write a prompt chain of thought

1. The zero-shot chain of thought: one instruction is enough

An instruction is added to the request requiring a detailed explanation of the reasoning before a conclusion is reached. Here is a typical office scenario, first without, then with a chain of thought.

before:

Our team of 12 is switching to working remotely 4 days out of 5. We rent 14 workstations at €650 per month. How many workstations should we keep and what will the annual savings be?

A quick model often responds with a single stroke, with a plausible figure but a hidden assumption (does everyone come on the same day? does no one come?).

after:

Our team of 12 is switching to working remotely 4 days out of 5. We rent 14 workstations at €650 per month each. How many workstations do we need to keep to cover 90% of the days we're physically present, and what annual savings does this represent? Work through the process step by step: first, state your assumptions about the distribution of attendance, then perform the calculations, and finally present the result in a table. Indicate any assumptions that need to be validated.

The answer outlines the assumptions (spread out attendance throughout the week or concentrated attendance), the number of people present per day, the number of workstations required, and then the cost savings. You can challenge a specific step instead of rejecting the entire result.

2. The chain of thought few-shot: show a solved example

One or more worked examples, along with the expected reasoning, are provided before the actual question. The model replicates the method. This is the original version by Wei et al., and the most effective when you want to impose a calculation order or an analysis framework.

Example. Invoice of €1,200 excluding VAT, 5% discount on the pre-tax amount, VAT 20%. Reasoning: discount = 1,200 × 5% = €60; net pre-tax amount = €1,140; VAT = 1,140 × 20% = €228; total including VAT = €1,368. Now it's your turn, using the same method: invoice of €3,450 excluding VAT, 8% discount on the pre-tax amount, VAT 20%.

The same principle applies to qualitative analysis: a single example of a customer complaint analyzed according to your framework (fact, impact, probable cause, proposed solution) is sufficient for the model to process subsequent complaints in the same order. For pre-structured business prompts with context, steps, and output format, see our Library of AI prompts categorized by profession.

Chain of thought, few-shot, prompt chaining, chain of verification: the differences

These terms often come together and are often confused. However, they refer to different levers, which are combined.

TechnicalPrincipleNumber of exchangesWhen to use it
Chain of thoughtWrite down the steps of the reasoning before the answer.Only oneCalculation, logic, multi-criteria decision-making, with a fast model
Few-shotShow examples of expected inputs and outputsOnly oneImposing a format, tone, or method; combinable with the chain of thought
Prompt chainingBreak down a task into several prompts; the completion of one feeds into the next.ManyA process in separable steps, each reviewed before the next
Chain of verificationHave students produce an answer, then answer control questions separately, and correct them.FreeFactual answers where the risk of fabrication (hallucination) is high
Self-consistencyGenerate several lines of reasoning and retain the most frequent answerManySingle-solution problems where reliability matters more than cost

Self-consistency was proposed in 2022 by Xuezhi Wang and his colleagues at Google, who showed that voting among multiple thought chains further improved the results of the simple chain of thought. The chain of verification was described in 2023 by a team at Meta (Dhuliawala et al.) to reduce hallucinations in factual responses.

Examples of prompt chaining: 6 prompt chaining examples to copy

Prompt chaining (also spelled "chain of prompts" or "chain prompting") doesn't require the model to reason more effectively; it breaks down the workload. Each prompt has a single objective, and you check the result before moving on to the next. An example of prompt chaining consists of three or four successive requests, each with a single objective. Example 1, based on customer feedback:

  1. Prompt 1: "Here are 40 customer comments. Classify them by theme and count the occurrences of each theme."
  2. Prompt 2: "Based on this ranking, identify the three main irritants and cite two representative verbatim quotes for each."
  3. Prompt 3: "Write a one-page memo for the management committee outlining these three irritants and providing one recommendation for each irritant."

If the ranking of prompt 1 is incorrect, you correct it before it contaminates the final grade. This is the purpose of chaining: to limit the propagation of errors. By 2026, agents and delegation tools of this type Cowork They apply this logic automatically: they break down a request into steps, chain the processes together, and only get back to you with the deliverable. Understanding prompt chaining helps to define the scope of what you entrust to them; see our guide Claude Cowork.

Five other chains, to be used as is in Copilot, ChatGPT, Claude or Gemini by replacing the documents with your own.

Example 2: Preparing a steering committee

  1. Prompt 1: "Here are the minutes from the last three project meetings and the schedule. List the milestones met, those that are late, and those that are still to come, with their dates."
  2. Prompt 2: "From this list, identify the three main risks and, for each, the decision to request from the committee."
  3. Prompt 3: "Draft a five-point agenda for the committee, specifying the time allotted for each point and the expected decision."

Example 3: Responding to a call for tenders

  1. Prompt 1: "Here is the specification document. Extract all the numbered requirements, distinguishing between mandatory and optional ones."
  2. Prompt 2: "For each requirement, indicate whether our attached offer meets it, partially or not at all, with the reference page."
  3. Prompt 3: "Write a summary of the response: strengths, gaps to address, questions to ask the buyer."

Example 4: Writing a job offer

  1. Prompt 1: "Here is the internal job description. Rephrase the responsibilities in six results-oriented sentences, without internal jargon."
  2. Prompt 2: "List the essential and desirable skills, removing any criteria that could be discriminatory."
  3. Prompt 3: "Write the advertisement in 250 words, in the style of our attached editorial guidelines, with a title and a catchy sentence."

Example 5: Analyzing a sales table

  1. Prompt 1: "Here is the export of monthly sales by region. Check the data: missing values, duplicates, incomplete months."
  2. Prompt 2: "Calculate the change in each region over twelve months and rank them from the largest increase to the largest decrease."
  3. Prompt 3: "Propose three possible explanations for the two regions showing a decline, and indicate what data would allow them to be verified."
  4. Prompt 4: "Write a ten-line commentary for the monthly report, separating facts from assumptions."

Example 6: Turning a meeting into an action plan

  1. Prompt 1: "Here is the transcript of the meeting. List the decisions made, along with the person who announced them."
  2. Prompt 2: "List the actions, each with a person responsible and a deadline; indicate those that have neither."
  3. Prompt 3: "Draft the summary email to the participants: the decisions first, then the action plan."

Three rules for proper chaining. One objective per prompt. A structured output, such as a numbered list or table, so that the next prompt can unambiguously build upon it. A review at each step before reusing the output. These chains are best learned using your own documents: this is the principle behind our learning paths. Mastering Generative AIwhere each participant builds their own chains for writing, summarizing, or analyzing. For the method of writing the instructions themselves, see our guide to prompt engineering.

The chain of verification: have the answer checked

The chain of verification addresses another problem: a well-reasoned answer can be based on a fabricated fact. The method consists of four steps, which can be written in a single prompt:

1. Answer the question: Which obligations under Section 50 of the AI Act apply to a user company? 2. Then list four verification questions that would allow you to check the statements in your answer. 3. Answer each question independently, without referring to your first answer. 4. Correct your initial answer if a verification contradicts it, and indicate what you have changed.

It does not replace human verification on a primary source, but it brings some errors to light before they appear in a document.

The chain of thought in the age of reasoning models

Since late 2024, developers have been offering models that generate their own internal thought process before responding: ChatGPT's "Reflection" mode, Claude's extended reflection, and Gemini's Thinking mode. In Copilot, as in other assistants, this often appears as an option or a dedicated model in the selector. The chain of thought hasn't disappeared; it has become a function of the model rather than a prompt trick.

The developers draw explicit recommendations from this. OpenAI advises against using "think step by step" prompts with its reasoning models, which already reason internally, and instead recommends keeping prompts simple and direct. Anthropic observes that a general instruction like "think deeply" often yields better reasoning than a detailed handwritten plan, because the model's reasoning frequently exceeds what a human would prescribe. Both developers, however, maintain the value of solved examples and numbered steps when order matters.

In practice, with a reasoning model, the work of the prompt shifts towards three elements:

  • The context : the data, constraints, and assumptions that the model cannot guess (your discount policy, your rounding rule, the population concerned). This is the purpose of the context engineering.
  • Success criteria: What constitutes a good answer, and what should lead to a solution being rejected.
  • The verification: request to check the result (recalculation, consistency of figures, sources) before delivering it, except with the most recent models which already do this automatically.

When to still use the explicit chain of thought

SituationShould we write "step by step"?What has to be done
Quick model (without thinking) on ​​a calculation or decisionYesRequest the steps, assumptions, and a final table
Reasoning model for an open problemNoProvide the context, objective, and success criteria
Mandatory method (analysis grid, regulatory calculation order)Yes, in the form of numbered steps or a solved example.Write the method, regardless of the model.
Sensitive factual responseNot very usefulChain of verification and source verification
A long-term task with multiple deliverables.NoPrompt chaining, or delegation to an agent
The result must be justified to a third party.YesAsk for the reasoning evident in the answer, so that you can review it.

To choose between a rapid model and a reasoning model depending on the task, see our article. The right AI model for each taskAnd to place chain of thought within the broader context of prompt writing techniques, our guide to prompt engineering resumes the complete method.

The limitations of the chain of thought

  • A plausible argument is not necessarily a correct argument. The model can go through convincing steps and make a mistake at one of them; each figure still needs to be recalculated.
  • The stated reasoning is not always the true one. Research on the fidelity of thought chains shows that the written explanation does not always reflect the calculation actually performed by the model.
  • More text means more cost and more time. A chain of thought lengthens the answer; on a simple question, it adds nothing.
  • A context error is propagating. If an input data is false, the steps faithfully reflect it until the conclusion.

That's why we teach chain of thought as a review habit as much as a prompting technique. In our workshops Prompt PartyParticipants compare the same request with and without a chain of thought, using their own cases, and then learn to identify the faulty step in a line of reasoning.

F.A.Q

What is the definition of a chain of thought?

The chain of thought is a prompting technique that prompts a language model to produce the intermediate steps of its reasoning before its final answer. It improves results on problems involving computation, logic, and decision-making.

What is the difference between chain of thought and prompt chaining?

The chain of thought structures reasoning within a single response. Prompt chaining breaks down a task into several successive prompts, the output of one serving as input for the next, with possible control between each step.

Should we still be writing "think step by step" in 2026?

With a fast model, yes, for calculation or logic tasks. With a reasoning model, no: it already reasons internally, and OpenAI and Anthropic recommend general instructions, context, and success criteria instead.

What is few-shot prompting?

This involves providing the model with a few examples of expected inputs and outputs before the actual demand. Combined with the chain of thought, each example also illustrates the reasoning to follow.

What is the chain of verification?

A four-step method described by Meta in 2023: answer, formulate verification questions, answer them independently, then correct the initial answer. It aims to reduce fabricated facts.

Does the chain of thought function work in Copilot or ChatGPT?

Yes, in all assistants. If you have a choice of modes, activate reflection for complex problems and provide context; in quick mode, explicitly ask for the steps.

What is a simple example of chaining prompts?

Analyze customer feedback in three steps: categorize the comments by theme, identify the three main pain points, and then write a one-page memo for the management committee. Each submission is reviewed before being used as input for the next step, preventing any classification errors from appearing in the final memo.

Sources

To go further: this topic resonates with our complete guide to prompt engineeringwhich also provides an overview of the profession of prompt engineer.

Back to top
All NextStart.AI offersEight consulting missions, seven pillars, around twenty formats, eighteen professions, five tools
Around thirty major organizations were supported.Large groups, public agencies and networks: from acculturation to production

Insurance and social protection

Public sector and research

Consumption and retail

The NextStart.AI collectiveConsultants, trainers and industry experts: more than 5 350 trained professionals since 2023
All free resourcesDiagnostics, Finders, prompt library, kits by profession, by tool, and for managing an AI program
The NextStart.AI blogGuides, prices and comparisons of Copilot, Claude, ChatGPT, Gemini and Mistral, and field analyses
Let's talk about your AI projectResponse within 24 business hours
AI agents in businessUnderstanding, choosing the tool, training the teams, deploying
Delegate entire tasksClaude Cowork, Copilot Cowork, ChatGPT Work, Vibe Work