In this article
- What if you do not know how to write a good prompt?
- What makes a marketing prompt good when everyone gives a role?
- Which prompt do you use as a strategist?
- Which prompts do you use as a creative?
- Which prompt do you use as a campaign builder?
- What mistakes does your team make with prompts?
- How do you build a shared prompt library?
- When does this not work?
- How do you practise prompting with your team?
- Sources
In short
- Not sure what to ask? Do not start with a prompt, start with a conversation: let the AI interview you, then write a prompt for you.
- A good marketing prompt does more than set a role and a task. It has the AI ask questions first, separates what you know from what the AI assumes, forbids invention and builds in a check.
- Below are an interview prompt and four prompts for three roles, with an explanation of why they work. Put the ones that work in a shared library.
Many marketers already use ChatGPT, Claude or Copilot and know the basics: give a role and context. So that is no longer where the difference lies. The difference is in what you make the AI do before it writes, and in what you forbid it.
Not so confident writing prompts? That is fine, because you do not have to do it alone. Describe where you want to get to and let the AI ask you questions. It writes the prompt from your answers. Then you do not have to guess what belongs in it.
What if you do not know how to write a good prompt?
Let the model interview you. Paste the prompt below, fill in one or two sentences on what you want to achieve, even if it is vague, and answer the questions.
You are an experienced prompt designer and marketing strategist. I want you to write a good prompt for this task: [describe in one or two sentences what you want to achieve, even if it is vague].
Interview me first. Ask questions in rounds of at most five and wait for my answer after each round. Ask, among other things, about: the goal and how we measure success, the audience, what we already have (texts, data, earlier results), what is not allowed, and who will use the result and where. For each question, say in half a sentence why you are asking.
Stop asking once you know enough, or after three rounds. Then write one prompt I can use straight away, structured as role, context, task and format. Put variables I need to fill in later in square brackets. Afterwards, list in at most three points which assumptions you made and what I should check.
Why it works. A model knows nothing about your brand, your audience or your earlier results. Asking questions is cheaper for the model than guessing, and cheaper for you than starting again three times. When it lists its assumptions, you see what to check.
The providers do not name this as a standard method. We suggest it as a way of working. A similar idea appears in Utrecht University's prompting guide, which mentions "reverse prompting": letting the AI write a prompt (Utrecht University). Anthropic says you can also ask Claude to evaluate your examples or generate extra ones (Anthropic).
Use this for work where a lot is at stake: a campaign, a strategy, a report. For a short email it is too much hassle.
What makes a marketing prompt good when everyone gives a role?
Role, context, task and format are the basis. Anthropic offers a test: give your prompt to a colleague who does not know the task and ask them to carry it out. If that colleague is confused, the model will be too (Anthropic). Google also advises clear instructions, context, examples and a format (Google). The four parts are our way of organising that advice.
Good prompts add five things on top:
- Ask first. The model asks its questions before it writes.
- Separate knowing from assuming. What comes from your input and what the model comes up with are kept apart.
- Invent nothing. Figures, names and results come from you. If there are none, none appear.
- Criticise itself. The model judges its own work against your goal before showing you anything.
- One decision. The outcome is a choice with a reason, not a list to pick from.
These are our own rules of thumb, not rules from the providers. You will see them in every prompt below. The explanation under "Why it works" is always our reasoning, not research.
Which prompt do you use as a strategist?
A strategist wants to justify choices, not be confirmed in what they already think.
You are a senior marketing strategist who thinks critically and does not just agree.
Context: we sell [product] to [audience]. Our goal this quarter is [goal]. This is what we already know: [three to five sentences, without names or personal data].
First ask at most five questions you need and wait for my answer. For each question, say why you are asking.
Task:
1. Formulate three hypotheses about which segments are most likely to respond. For each hypothesis, separate: what we know (from my context), what you assume and which finding would disprove the hypothesis.
2. For each hypothesis, give a four-week test with a KPI and the result at which we stop.
3. Then play the opponent: why is the hypothesis you find strongest likely to fail?
Format: two tables. Table 1: hypothesis, what we know, what you assume, what would disprove it. Table 2: hypothesis, test, KPI, stopping point. Then at most 150 words for step 3. End with one recommendation: which test first and why.
Why it works. A hypothesis with a disproving finding forces the model to be honest about what it does not know. The opponent role is meant to take the nodding along out: if you only ask a model for advice, you mostly get arguments for that advice. That is our reasoning and not researched fact, so judge for yourself whether the pushback adds anything.
Which prompts do you use as a creative?
A creative wants copy in the brand's own tone. Do not describe that tone yourself, have it drawn from your best texts. Good and bad examples together are strongest. Google advises examples in your prompt (Google) and Anthropic calls examples one of the most reliable ways to steer tone and structure (Anthropic).
Step 1 builds the tone profile, which you then keep:
You are a brand editor. Below are five texts from [brand] that work well: [texts]. And two texts we do not like: [texts].
Task: describe our tone in a profile of at most 200 words: sentence length, word choice, how we address the reader and what we never say. For each trait, quote a sentence from the examples as evidence. If the examples contradict each other, say so.
Step 2 uses that profile for the hooks:
You are a copywriter who writes for [brand]. Use this tone profile: [profile].
Context: the audience is [audience], the goal is [goal] and the topic is [topic]. Figures you may use: [real figures, or none].
Task:
1. Write fifteen hooks for a LinkedIn post, three per angle: question, figure, claim, anecdote and counter-example.
2. Judge your own hooks strictly on three points: does it fit the tone profile, does it serve the goal, and could it apply to more than one company? Cut whatever scores badly.
3. Pick the best three and explain in one sentence why for each.
Do not invent figures. An anecdote is fictional unless I give it above: write it as "Imagine" or leave it out. Put an asterisk next to every claim I should verify.
Format: first the three chosen hooks. Then the other twelve hooks, each with one short line on why you cut it.
Why it works. The profile with evidence sentences shows whether the model really understands your tone. "Could it apply to more than one company?" removes the generic hooks. Writing fifteen and picking three gives you a wider range than five, and the model ranks them instead of you reading everything. How to build a role for social posts with AIDA and PAS is in The anatomy of a good prompt: the copywriter.
Which prompt do you use as a campaign builder?
A campaign builder wants to learn, not just publish. So you design an experiment, not four separate ads.
You are a performance marketer who designs experiments for [channel].
Context: we advertise [product] to [audience]. The goal is [goal]. The audience's main objections are [two or three objections]. The channel has these character limits: [limits]. Proof we may mention: [real figures or a customer quote, or leave empty]. Budget per week: [amount], expected clicks per week: [number].
Task:
1. Formulate one hypothesis about what holds the audience back.
2. Write four variants that change one variable, only the angle, within the character limits.
3. Do not invent figures, customer names or results. If I have given no proof, do not mention a figure.
4. Say how long and at how many clicks we only draw a conclusion, and when we stop.
5. Say which result would disprove the hypothesis.
Format: a table with variant, angle, headline and text. Below it, the measurement plan in at most five sentences.
Why it works. One variable per variant makes the outcome readable: afterwards you know what worked. A stopping point and a disproving result stop you running a test until you see something nice. For the proof angle the model needs real proof from you, otherwise it may not mention a figure.
What mistakes does your team make with prompts?
- Too little context. "Write a newsletter" makes the model guess. Google names context as one of five tips for Gemini (Google Workspace).
- Everything in one prompt. A strategy, five texts and a plan in one request gives a flat result. Google advises splitting complex tasks across separate prompts.
- Not checking the output. Check figures, sources and claims in AI text yourself before you publish them.
- Customer data in a public tool. Do not put personal data in a tool your organisation has not approved, not even in an interview or a data export. How to agree that is in Using AI safely: how to write an AI policy for your team.
- Everyone writing their own prompt. Then every colleague gets a different tone and quality, and a good prompt is lost when someone leaves.
How do you build a shared prompt library?
- Start small. Collect the five prompts your team uses every week.
- Put them in the same structure. Role, context, task and format, with the variables in square brackets.
- Record who owns each prompt. That is the person who maintains and replaces it.
- Note when you last tested it and with which tool. Anthropic warns that its techniques were measured on a particular model and that you should re-check them when you move to another. OpenAI advises building evaluations so you can measure behaviour when models update (OpenAI).
- Remove what nobody uses. Nobody uses a library of a hundred prompts.
A line in your library can look like this (fictional example):
| Prompt | Owner | Last tested |
|---|---|---|
| Strategist: hypotheses and opponent | Marketing manager | October 2026, Claude |
When does this not work?
A prompt does not replace strategy or data. Without audience insight, the model writes a plausible story that rests on nothing. A prompt also does not give an identical result every time: for a fixed, frequently repeated task, a dedicated assistant with fixed instructions is better than a prompt you paste each time. And what works for one model does not always work for another.
The interview takes time. For a short, clear task an ordinary prompt is quicker. Save the interview for work where a wrong assumption costs you days.
We tried these prompts out with a fictional company, in one model. Test them in the tool your team uses before you put them in your library. Look above all at whether the model sticks to "invent nothing" and whether the figures add up.
How do you practise prompting with your team?
In the workshop AI for marketers, your team works through the roles of strategist, creative and campaign builder, with your own brand as the case. Participants take home a prompt library per role.
In January 2025 we published a list of 25 prompts for each stage of the B2B funnel. You can still download it (in Dutch). Rework them into the four parts above and they become more useful.
Sources
- Anthropic, Prompting best practices, accessed 23 October 2026
- Google AI for Developers, Prompt design strategies, updated 13 October 2026
- Google Workspace Help, five tips for writing prompts for Gemini, accessed 23 October 2026
- OpenAI, Prompt engineering, accessed 23 October 2026
- Utrecht University, Artificial Intelligence (AI): Prompting, 8 September 2026
