Prompt engineering: how to write an effective AI prompt

What is prompt engineering?

Prompt engineering is the practice of writing, testing and adjusting the instructions given to a generative AI model. It is the core skill of AI marketing: every use, from email subject lines to product descriptions, goes through an instruction.

The same prompt can give two different answers: OpenAI calls prompting “a mix of art and science”.

Writing a prompt is not the same as optimizing to be cited by ChatGPT: that work belongs to GEO, as does answer engine optimization. In GEO, a prompt panel serves a different purpose: measuring a brand’s visibility, with fixed questions recorded at set intervals, not working instructions.

What is an AI prompt?

An AI prompt is the question or instruction addressed to a model. It is often longer and more precise than a query typed into a search engine, and it can contain context, documents and examples.

The instruction is not limited to the message you type: OpenAI describes system instructions that set the identity, rules, examples and context in advance. Claude Projects (five on the free plan) and Gemini Gems store reusable instructions.

What do Anthropic, OpenAI and Google recommend?

The three vendors agree on the basics: a clear instruction, context, examples, a specified format and repeated testing. Anthropic’s golden rule: “Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they’d be confused, Claude will be too.” OpenAI compares a reasoning model to “a senior co-worker” you give a goal to, and a GPT model to “a junior coworker” who needs explicit instructions.

Pages accessed on October 19, 2026; Gemini API guide updated on October 13, 2026; Google Workspace guide from December 2025.

Two disagreements stand out: the number of examples, and roles, on which Anthropic’s documentation and its best practices article contradict each other. We do not settle the question: each guide describes its own vendor’s models, at a given date.

How do you write an effective prompt for marketing?

Write your prompt like a brief for a contractor who knows neither your brand nor your customers: objective, context, audience, tone, constraints, allowed sources and things to avoid. This is the method Elev8 Lab recommends. The Google Workspace guide keeps four components, Persona, Task, Context and Format, without requiring you to use all of them.

1. Outcome: describe a good output and how to judge it (Anthropic’s success criteria).
2. Context: describe the audience using your persona sheet and paste the rules of your editorial line rather than a tone adjective. Say why each instruction matters.
3. Task: phrase it with a verb (write, sort, summarize), which Google Workspace calls “the most important part of a prompt”.
4. Sources: provide product sheets, notes or approved figures, and ask the model to stick to them.
5. Format: set length and structure; write “write in paragraphs” rather than “no lists”.
6. Example: add a piece of your brand’s content that has already been used.
7. Uncertainty: allow the model to say it does not know.
8. Iteration: rephrase, change the order of the elements or split a complex task into several chained prompts.

Example: from a vague request to a complete brief

Illustrative example, built for this article: it does not come from any real campaign and is not linked to any result. Square brackets mark the variables.

Vague version: Write email subject lines to follow up on abandoned carts.

Complete version, in six blocks to paste in order:

• Objective: “Suggest 5 subject lines for an abandoned cart follow-up email. A good subject line makes people want to open the email without promising a discount.”
• Context: “We sell [product type] to [description of the audience]. Our tone: [three rules from the editorial line]. The email goes out [delay] after the cart is abandoned.”
• Constraints: “[Number] characters maximum per subject line. No discount is planned: do not promise one.”
• Examples: “Here are two subject lines we have already sent, in our tone: [subject line 1]; [subject line 2].”
• Format: “Return a table with three columns: subject line, angle chosen, character count.”
• Missing information: “If you are missing any information needed to follow these instructions, ask me before you start writing.”

Each block settles a question the model would otherwise decide on its own. These subject lines then go into your email marketing campaign.

Which prompting techniques actually work?

Three families of techniques recur in the guides: zero-shot (an instruction with no example) or few-shot (a few examples), step-by-step reasoning, and roles. Tests with a published method put the last two into perspective.

These tests, run by the Generative AI Labs at the Wharton School (University of Pennsylvania), use difficult multiple-choice questions (GPQA Diamond, MMLU-Pro). They measure accuracy, not the quality of a marketing text.

Should you give the AI an expert role?

Not to gain accuracy. Across six models tested, Prompting Science Report 4 (Basil et al., December 5, 2025) concludes that expert personas generally did not improve accuracy compared with a prompt without a role; only Gemini 2.0 Flash is an exception. A role outside the question’s domain sometimes degrades the answers, and a low-knowledge role (“young child”) often does.

Politeness, tips, threats: do they work?

None has a guaranteed effect. Being polite to the model sometimes improves performance and sometimes lowers it, depending on the question (Prompting Science Report 1, March 4, 2025). Promising a tip or threatening the model generally has no significant effect (Prompting Science Report 3, August 1, 2025); this report tested a claim made by Sergey Brin in May 2025.

“Think step by step” helps models without built-in reasoning a little, on average, but makes answers more variable: some questions answered correctly without the phrase fail with it (Prompting Science Report 2, June 8, 2025).

How do you check what the AI produces?

Review every output, ground it in sources you provide and remove any claim the model cannot support. When the task is finding a source, measured errors are frequent.

Anthropic describes three techniques on its Reduce hallucinations page: allow the model to say “I don’t know”; extract exact quotes from a long document before answering; back each claim with a quote and remove the others, for example in a press release drawn from product sheets. These techniques reduce hallucinations without eliminating them.

For recent or little-known facts, Google recommends the Gemini API’s Grounding with Google Search. In France, the CNIL (the data protection authority) advises checking outputs for accuracy, plagiarism and stereotypes, and never reproducing an output as is; for Google Workspace, “the final output is yours”.

Style needs reviewing too: OpenAI’s guide to the GPT-6 model (accessed on October 19, 2026) provides a prompt against stock phrases such as “delve” or “leverage”. For content meant for search, Google’s position is detailed in our guide to SEO writing with AI.

What data should you never put in a prompt?

Only enter information you are authorized to share. The CNIL recommends never entering personal data, or company data covered by secrecy such as trade secrets, into a consumer AI service (CNIL questions and answers, July 18, 2024).

A second risk: prompt injection, defined in our SEO and GEO glossary. OWASP ranks it first among the risks of large language model applications (LLM01:2025). In its indirect form, instructions hidden in a web page or a file read by the AI take effect, even when they are invisible to a human. OWASP’s countermeasures: flag untrusted external content, limit the tool’s permissions, require human approval for high-risk actions.

The broader legal framework (AI Act, GDPR) is covered in our AI marketing guide.

How do you build a prompt library for your team?

Start from the tasks your team repeats, write a template prompt for each, test it, then date it. The CNIL itself suggests that organizations provide “prompts types” (template prompts) for identified, proven tasks.

1. Tasks: choose recurring requests (email subject lines, SEM ads, product descriptions).
2. Record: note the objective, the variables in square brackets, an example of good output, and the date and model of the last test.
3. Test: try the prompt on several real cases before sharing it.
4. Sharing: save it in the instructions of a Claude Project (shared within the organization on Team and Enterprise plans) or in a Gemini Gem.
5. Review: retest it whenever the model changes; according to OpenAI, two versions from the same model family can produce different results.

Model names change fast, and OpenAI recommends testing production prompts before every change, like code. For a marketing team, the equivalent we suggest: a dated record and a new test before any update.

Prompt engineer: career of the future or basic skill?

For a marketer, prompt engineering is a basic skill. The Google Workspace guide says so: “You don’t have to be a prompt engineer to use generative AI.”

Anthropic presents “context engineering” as “the natural progression of prompt engineering” (Effective context engineering for AI agents, September 29, 2025). The work becomes “less about finding the right words and phrases for your prompts” and more about choosing the whole set of information given to the model: instructions, history, documents, tools. This is especially true for AI agents.

To learn: Anthropic’s free interactive tutorial (written for Claude 3, to be supplemented with its current documentation) and the Marketing section of the Google Workspace guide. We found no primary source on salaries or on the number of prompt engineer job offers in France.