Prompt engineering: the method for writing effective prompts (2026)

This article was published on December 8, 2025, and rewritten and updated on October 24, 2026, based on the official prompt engineering guides from Anthropic, OpenAI, Google, and Microsoft, which were consulted that day. The examples are taken from our in-house workshops.

Prompt engineering is the art of formulating instructions for a generative AI (ChatGPT, Copilot, Claude, Gemini, Vibe de Mistral) to obtain a correct, useful, and well-formatted response on the first try. In practical terms, it involves telling the AI ​​what it should be, what it should do, based on what information, for whom, and in what format, then testing and improving the formulation. This guide provides the structure of a good prompt, the techniques that really work in 2026, the mistakes to avoid, and a method for building reliable prompts in a team.

To see these principles applied, the free NextStart library offers ready-to-copy prompts, such as the one for moving from a blank page to a first deliverable.

The essentials in five lines

• A good prompt combines five components: a role, a task, a context, sources, and an expected output format.
• The golden rule Common to all editors: be clear and direct, as with a colleague who knows nothing about the case.
• The techniques that matter: Explain the why, give examples, separate the parts of the prompt, break down complex tasks, ask for verification.
• What has changed with the reasoning models: There's no need to dictate "step by step" to them; it's better to precisely describe the goal and the success criteria.
• In company, A prompt that is frequently used is tested on several cases, researched and shared, instead of remaining in the head of its author.

What is prompt engineering?

A prompt is the text you give to a language model: a question, an instruction, a document to process. The model doesn't read your mind; it responds to what is written. Prompt engineering, therefore, consists of reducing the gap between what you want and what you have actually asked for.

The term originated in the developer world, where prompts are written for applications. But today it applies to all users: Google's guide for Gemini in Workspace notes that the most effective prompts average around twenty words, while most users write fewer than nine. This discrepancy explains a good deal of disappointment: "summarize this document" lets AI guess the length, the target audience, the key takeaways, and the format.

The simplest test comes from Anthropic's documentation: show your prompt to a colleague who is unfamiliar with the topic. If they get confused, the AI ​​will too.

The structure of a good prompt

Each publisher offers its own framework. Microsoft discusses objective, context, sources, and expectations for Copilot; Google uses persona, task, context, and format for Gemini. These frameworks overlap into five components, not all of which are necessary in every case.

Component	The question she answers	Example
Role	What should AI be?	"You are a management controller in an industrial group."
Task	What should it produce, and with which action verb?	"Write a report analyzing the budget variances."
Context	Why, for whom, and in what situation?	"The memo will be read by the management committee on Monday, which must decide whether or not to freeze two projects."
Sources	Based on what information?	"Use only the attached Budget_T3.xlsx file; report any missing data."
Format	In what form, what length, what tone?	"One page maximum: three key findings, two recommendations, a summary table, and a factual tone."

Taken together, these elements produce a prompt of about sixty words, which generates a usable note instead of a generic summary. The most frequently overlooked component is context: yet it is this that allows the AI ​​to choose what matters.

Techniques that work, with an example for each

1. Be clear and direct

Recent models follow the guidelines to the letter: if you want more than the bare minimum, you have to ask for it. Vague verbs ("look," "help me with") produce vague answers.

• before: "Help me with this email from a dissatisfied customer."
• after: "Write a response to this email from a dissatisfied customer: acknowledge the delivery delay, offer a 10% discount, give a firm delivery date, in fewer than 120 words and in a courteous tone."

2. Explain the reason why

Anthropic recommends giving the reason for an instruction rather than the instruction itself: the model generalizes from the explanation. A justified rule is better followed than a prohibition written in capital letters.

• before: "NEVER use jargon."
• after: "This text will be posted on the intranet and read by field agents who are not specialists: replace each technical term with a simple explanation."

3. Give a few examples

An example is worth a thousand words. Anthropic's documentation recommends three to five relevant and varied examples so that the model grasps the principle without copying a single detail. With reasoning models, OpenAI, on the other hand, advises starting without examples and adding them only if the result is unsatisfactory.

• before: "Organize these customer comments by theme."
• after: "Classify these customer comments into one of the following themes: Delivery, Price, Quality, Service. Examples: 'Package arrived three days late' → Delivery; 'The advisor never called me back' → Service; 'Cheaper at the competitor' → Price."

4. Separate the parts of the prompt

When instructions, documents, and data are quickly mixed together, AI sometimes confuses them. Clear delimiters (titles, quotation marks, tags like And The approach recommended by Anthropic and OpenAI prevents a sentence in the document from being interpreted as an instruction. For long documents, Anthropic advises placing the text first and the question at the end.

• before: a contract stuck in the middle of three lines of instructions.
• after: " [contract text] "Based on the contract above, list the termination clauses, citing the exact article for each."

5. Impose the output format

A precise format makes the answer immediately usable: a table with named columns, a numbered list, a word count, or a prescribed structure. It's better to say what you want than what you don't want: "write in continuous paragraphs" works better than "don't use bullet points."

• before: "Compare these three offers from service providers."
• after: "Compare these three offers in a table with columns for Total Price (excluding VAT), Delivery Time, Guarantees, and Points to Consider, then recommend one in three sentences."

6. Break down complex tasks

A single prompt that asks for analysis, recommendations, and drafting all at once yields an average result across the three. It's better to chain together short prompts and check each step: this is called prompt chaining. Anthropic's most useful approach is to produce a draft, have the AI ​​review it based on predefined criteria, and then have it corrected.

• before: "Analyze this satisfaction questionnaire and draft the action plan."
• after: three successive prompts: "Identify the five most frequently cited irritants, with their frequency", then "For each one, propose two actions ranked by cost", then "Write the action plan based on the selected actions".

7. Make them reason, then check

Asking a classical model to explain its reasoning before concluding improves the answers to multi-step problems: this is the chain of thought promptingAdding a final verification instruction ("before answering, check that each number comes from the file") catches some of the errors.

• before: "Which supplier is the cheapest over three years?"
• after: "Calculate the total cost over three years for each supplier, showing your calculations, then indicate the cheapest option. Before concluding, check that you have included the activation fees."

What changes with reasoning models

Since 2025, most assistants have offered models that think before responding: "reflection" or "deep reasoning" modes from ChatGPT, Claude, Gemini, Copilot, or Vibe. They change three habits.

• No need to dictate the steps anymore. OpenAI advises against asking its reasoning models to "think step by step": they already do this, and the instruction can degrade the result. Anthropic observes that a general instruction ("think deeply") often works better than a detailed handwritten plan.
• The goal and the criteria matter more. What makes the difference is the precise description of the expected result and what makes it acceptable.
• Fewer defensive instructions. Recent models follow instructions better; warnings in capital letters inherited from older models sometimes cause excessive caution.

Prompt engineering does not disappear: it shifts from formulation tricks to clarity of need.

Prompt engineering or context engineering?

With agents and assistants connected to your documents, a growing portion of what the model reads no longer comes from the prompt: attachments, search results, conversation history, standing instructions, tools. Anthropic describes context engineering as the evolution of prompt engineering: the set of strategies for selecting and maintaining the most useful information in the model's window.

	Prompt engineering	Context engineering
Subject	The formulation of the request	Everything the model has in front of it when responding
Who is in charge of it?	Each user	The designer of an agent, an assistant, or a project
Levers	Role, task, context, examples, format	Reference documents, standing instructions, memorandum, tools, source selection
Example	Be sure to request a meeting summary	Build an agent that can independently find the correct report and figures

The two complement each other: a well-fed agent remains disappointing if their instructions are unclear. We detail this evolution in From prompt to context: why context engineering is replacing prompt engineering.

The most common mistakes

• The prompt telegraph. "Summarizes" or "improves" without an audience, without length, without an objective: AI fills the gaps with average choices.
• To ask for everything at once. Analysis, recommendation and writing in the same prompt: each part is treated superficially.
• The conflicting instructions. "Be exhaustive" and "in five lines": the referee model for you.
• Forget the sources. Without a reference document, the AI ​​responds with its general knowledge, sometimes outdated or invented.
• Do not reread. A good prompt reduces errors without eliminating them: figures, names and quotations are always verified.
• Keep your best prompts to yourself. A quick win that works and that no one else knows about is a lost opportunity for the team.

Building and testing a prompt in a company: the method

For a one-off question, a well-structured prompt is sufficient. For a task that recurs weekly within a team (reporting, analyzing complaints, preparing for meetings), the prompt becomes a working tool, deserving the same rigor as a document template.

1. Start with a real-life case. Take a real case from the team and the result that an expert would have produced by hand: that is your reference.
2. Write the first draft with the five components from the table above.
3. Test on three to five different cases. including a difficult or atypical case. A prompt that only works on the example that inspired it is not reliable.
4. Compare to the reference result with simple criteria: accuracy, completeness, format, time required for correction.
5. Correct one thing at a time, to find out what improved or worsened the result.
6. Document and share. The final prompt, its use, its limits and the checks to be made go into a common library, or become the instructions of an agent or a shared project.
7. Retest after each model change. The techniques vary from one model to another: a perfect prompt on one version may behave differently on the next.

This is the method we use in our workshops. Prompt Party In one hour or half a day, each participant builds and tests their prompts on their own projects, using the tool chosen by their IT department. To get started without a blank page, our AI prompt library by profession gathers prompts ready to copy into Copilot, ChatGPT, Claude or Mistral, and our article How to improve your generative AI prompts? reverts to basic reflexes. To see what users of these tools actually do, the analysis of 1,5 million ChatGPT conversations gives a good overview.

F.A.Q

Is prompt engineering still useful in 2026?

Yes, but it has changed in nature. Modern models are better at understanding imperfect requests and reasoning independently, so clever wording matters less. However, accurately describing the purpose, context, sources, and expected outcome remains what distinguishes a generic response from a usable one.

What is the difference between a prompt and prompt engineering?

The prompt is the text sent to the AI. Prompt engineering is the process of designing this text, testing it on several cases, and improving it until a reliable result is obtained.

Do you need to know how to code to do prompt engineering?

No. For use in ChatGPT, Copilot, Claude, or Gemini, everything is done in plain language. Code only comes into play when developers integrate prompts into an application or automate their evaluation.

Does the same prompt work on ChatGPT, Copilot and Claude?

The principles are the same everywhere, and a well-structured prompt yields good results on all assistants. The details vary: Copilot can retrieve your emails and Microsoft 365 files, Claude makes good use of structure tags, and reasoning models prefer instructions without prescribed steps. Test an important prompt on the tool where it will be used.

How many examples should be given in a prompt?

For a classic model, three to five varied examples are generally sufficient, according to Anthropic's documentation. For a reasoning model, start without examples and only add them if the result is unsatisfactory, as advised by OpenAI.

Does a longer prompt always give a better result?

No. What matters is that each sentence provides useful information: the purpose, the context, a constraint, an example. A long but confusing or contradictory prompt is less effective than a short and precise one.

Sources

• Anthropic, Prompting best practices, Claude documentation, accessed on October 24, 2026.
• Anthropic, Effective context engineering for AI agents, September 29, 2025.
• Microsoft Learn, Write an effective prompt (Microsoft 365 Copilot).
• Google Gemini for Google Workspace, Prompting guide 101.
• Analytics Vidhya OpenAI's advice for prompting reasoning models.