The prompt formula: six building blocks
The prompting guides from OpenAI, Anthropic and Google say essentially the same thing: clear instructions, enough context, a defined output format and examples. For marketing work, this order has worked best for us:
- Role
Who should answer? “You are an experienced B2B newsletter copywriter.” That sets vocabulary and standards.
- Context
Brand brief, audience, offer, past results. The most important block and the one most often missing.
- Task
One clear verb and one outcome: “Write 12 subject lines”, not “Help me with my newsletter”.
- Format
Length, character count, table or list, number of variants, so you can use the output directly.
- Examples
Two or three of your own pieces that worked. Examples steer voice and quality more than any description.
- Limits
What must not happen: no invented numbers, no promises, max X words, [check] where information is missing.
You don’t have to write all six every time. Role, context, examples and limits live in the brand brief you set up once. Per task you only add the task and the format. That’s why many templates below start with [Brand brief].
Three rules that improve almost any prompt
- Ask for variants. Get 10 options instead of 1 and pick. Judging is easier than writing.
- Ask it to ask you. “Ask me up to three questions before you start” stops the AI from filling gaps with assumptions.
- Make it flag gaps. “If something is missing, write [check]” is the most effective guard against made-up numbers.
The foundation: your brand brief as a prompt
Set this brief up once and store it where your assistant always has it: as a project or custom instructions in ChatGPT, as project knowledge in Claude, as a Gem in Gemini. After that, the placeholder [Brand brief] in the templates is all you need.
You are the head of marketing at [company]. Remember this brand brief and follow it in every answer that follows. OFFER: [What we sell, in one sentence, with a price range] AUDIENCE: [role, age, situation] – problem: [their biggest problem] – buying trigger: [the moment they start looking] BENEFITS: 1) [benefit + proof] 2) [benefit + proof] 3) [benefit + proof] VOICE: [3 adjectives, e.g. direct, warm, no hype]. How we do NOT sound: [example sentence] NO-GOS: no income or health promises, no invented numbers, testimonials or sources. If something is missing, write [check]. WORD LIST: We say [term], never [term]. EXAMPLES: [2–3 of our own pieces that performed well] Briefly confirm you understood the brief and ask me up to three questions if anything is missing.
Strategy and positioning (1–5)
For the decisions that shape everything else. AI gives you options and counter-arguments; the call is yours.
Here is our offer: [description]. Write 10 positioning statements in the pattern “For [audience] who [problem], [offer] is the [category] that [benefit], unlike [alternative].” Score each 1–5 for clarity and differentiation and explain the top three.
Here is copy from three competitors (homepage, offer): [copy]. For each, summarise: promise, audience, price anchor, tone, proof. Then show in a table which customer problems none of them addresses. Don’t invent anything that isn’t in the text.
Create a SWOT analysis for [offer] in [market] based on these facts: [facts]. Then play devil’s advocate: which three assumptions in my strategy are most likely wrong, and how could I test each cheaply within a week?
Our core offer: [offer, price]. Design a four-step offer ladder: free entry point, entry offer under [amount], core offer, premium. For each step give content, price range, the move to the next step and the risk that it cannibalises the core offer.
Our goal for next quarter: [goal]. Starting point: [current numbers]. Derive three measurable marketing goals, each with a metric, baseline, target, weekly measurement and an early warning sign that tells me after four weeks it isn’t working.
Audience and personas (6–10)
AI helps you find patterns in customer feedback you already have. It doesn’t replace talking to customers.
Here are [number] anonymised customer comments: [text]. Build 2–3 personas with: situation, goal, biggest frustration, words they use themselves (quoted verbatim), objections and buying trigger. Tag anything you can’t derive directly from the text as [assumption].
Our product: [product]. Write 8 jobs to be done in the pattern “When [situation], I want to [motivation], so I can [outcome].” Rank them by frequency and urgency and suggest one interview question for each of the top 3 to validate it with real customers.
I’m running 30-minute interviews with [customer type] about [topic]. Create a guide with 10 open questions about past behaviour, not opinions. Add a follow-up for each and flag questions likely to produce polite rather than honest answers.
Audience: [audience]. Offer: [offer, price]. List the 10 most likely objections before buying, grouped by price, trust, effort and timing. Write an honest answer to each in two sentences max, with no pressure and no promises we can’t keep.
Here are forum posts, reviews or comments from our audience: [text]. Collect, verbatim, the phrases they use to describe their problem and desired outcome. Put them in a table: phrase, frequency, emotion behind it, where we could use it (headline, ad, subject line).
Ads (16–20)
AI is great at variants. The test decides what works, not the AI.
[Brand brief]. Offer: [offer]. Write 20 variants of a Meta ad: primary text (max 125 characters), headline (max 40 characters). Four per angle: problem, outcome, objection, proof, curiosity. No income claims, no superlatives without proof.
Keyword: [keyword]. Work out the search intent (learn, compare, buy). Then write 15 headlines (max 30 characters) and 4 descriptions (max 90 characters) that match that intent exactly. Check every character count at the end and flag any that run over.
Write a creative brief for a 20-second video for [platform]. Goal: [goal]. Content: hook in the first 2 seconds, 3 scenes with visual, on-screen text and voiceover, ending with a call to action. Note whether AI actors appear and how that’s disclosed.
Here are my ad results as a table: [data with impressions, clicks, cost, conversions per ad and placement]. What stands out? Which ads should I stop, keep or test? Back it up with the numbers and say explicitly where the data is still too thin to decide.
Check this ad copy: [copy]. Flag anything that could read as an income or health claim, an unsupported superlative, misleading pricing or a missing ad disclosure. Suggest a safer wording for each. This is a first pass, not legal advice.
Email and newsletters (21–25)
Your newsletter is a channel you own. Care beats volume here.
[Brand brief]. Lead magnet: [title]. Draft a 5-email welcome sequence over 10 days. For each email: goal, subject line (2 options), preview text, outline in 5 bullets, one single link. No product recommendation before email 3, and label it as an ad when it comes.
Here’s my newsletter: [text]. Write 12 subject lines (max 45 characters) in four styles: benefit, curiosity, question, number (only numbers from the text). Add preview text for each (max 80 characters) that complements rather than repeats the subject. No all-caps words, no spam triggers.
Here are my notes and links from this week: [notes]. Turn them into a newsletter of max 400 words: a personal two-sentence opener, 3 sections with one insight each, one practical tip, and a closing question for readers. Keep my wording where it works.
I recommend [product] (affiliate link, I earn a commission). Write an email with one goal and one link: the problem, how I solve it myself, who the product suits and who it doesn’t, the price with the date checked, and “Ad · affiliate link” right next to the button. No fake scarcity.
Write a short re-engagement email to subscribers who haven’t opened in a while. Tone: honest, no guilt-tripping. Content: what they missed (2 highlights: [topics]), asking whether they still want emails, one click to stay, one click to unsubscribe. Add three subject lines.
SEO and GEO (26–30)
For pages that get found in Google and in AI answers. The expertise and sources have to come from you.
Topic: [topic]. Main question searchers have: [question]. Create an outline with H2 and H3 headings that answers the main question in the first two sentences, then covers the typical follow-up questions. Mark where first-hand experience, sourced figures or an example are needed rather than making them up.
Here’s my page copy: [text]. Write 8 real questions readers would ask about it and answer each in 2–3 sentences using only information from the text. If the text doesn’t answer a question, write “missing from the text” so I can fill the gap.
Here’s my page: [text]. Write a short answer to the question [question] in two sentences max that makes sense without context and can be quoted, including the key fact and the caveat. Then give three variants with a different focus.
Page: [topic, main keyword, audience]. Write 5 title tags (max 60 characters including the brand) and 5 meta descriptions (max 155 characters) that name a concrete benefit. Count the characters and show them. No clickbait.
Here are the headings of our guides: [list]. Here are our customers’ questions: [list]. Which questions don’t we answer anywhere yet? Suggest 10 new pages, each with a working title, main question, audience and which existing page should link to it.
Sales and leads (31–35)
For enquiries and conversations. Cold outreach rules vary by market; in Germany, cold marketing email without consent is illegal.
[Brand brief]. Suggest 10 lead magnets that lead straight to our offer [offer]. For each: format (checklist, calculator, template, audit), the user’s result in one sentence, our effort (1–5), and why an AI assistant couldn’t produce that result in seconds.
Here is an enquiry: [text, without names or contact details]. Write a reply that answers the question directly, asks two targeted follow-up questions and proposes a concrete next step (e.g. a 20-minute call with two time options). Max 150 words, friendly, no filler.
I have a discovery call with a [role] at a [industry, size] company. Public info: [info]. Prepare: 3 hypotheses about their problem, 8 questions to test them, 3 objections with answers, and how to end the call with a clear next step.
Here are my call notes: [notes without personal data]. Write a summary for the client: their situation in their words, goal, proposed approach in 3 steps, what we need from them, timeline, next step. Don’t invent services we didn’t discuss.
Write three short follow-ups for a prospect who hasn’t replied to our proposal for [days] days: after 3 days (a useful addition), after 10 days (one specific question), after 21 days (a friendly close). 60 words max each, no artificial urgency.
Video scripts (36–40)
For reels, TikToks, Shorts and explainers. Idea and hook first, tech second.
Topic: [topic]. Write a 30-second reel script: hook (0–2 s), problem (2–8 s), three steps (8–25 s), close with a call to action (25–30 s). For each part give voiceover, on-screen text and a visual idea. Write the voiceover the way people talk, not the way they write.
Topic: [topic, e.g. AI notetakers in meetings]. Develop 5 ideas for a 20-second sketch based on an everyday office moment, two people max, with a punchline at the end. For each: the situation in one sentence, dialogue, punchline. No jokes at the expense of groups of people.
Explain [topic] to [audience] in a script for a 2.5-minute video: open with a question, three chapters with one key point and an example each, a short recap, and a pointer to [further resource]. Mark where a graphic or screenshot would help.
Here’s the script or transcript of my video: [text]. Create versions for TikTok, Instagram Reels, YouTube Shorts, LinkedIn and a newsletter paragraph: title or caption, length, editing note and call to action in each platform’s style.
Here’s my video transcript with timestamps: [transcript]. Fix obvious transcription errors without changing the content, trim filler words for subtitles, and suggest 5–7 YouTube chapters with timestamp and title (max 40 characters).
Analysis and reporting (41–45)
AI spots patterns fast. Check the maths with your own spot checks.
Here are my marketing numbers for this month and last: [table with visitors, sign-ups, open and click rate, cost, revenue]. Calculate the changes, name the three most important patterns with a possible cause, and suggest one test per pattern for next month. Show your working.
Here are our survey responses (anonymised): [data]. Summarise closed questions as a table with percentages and group open answers into 5–7 themes with frequency and one verbatim quote each. Flag where the sample is too small to support a claim.
Our funnel: [steps with numbers, e.g. visitors → sign-up → confirmation → click → purchase]. Calculate the rate at each step, compare it with typical ranges you clearly label as estimates, and name the step with the biggest lever. Suggest three actions sorted by effort.
Here are cost and results by channel for last quarter: [data]. Calculate cost per lead and per customer where possible. What would you do with [amount] of extra budget, and what should I measure first to decide with more confidence?
I want to test: [idea]. Turn it into a testable hypothesis (“If we change X, Y rises by at least Z, because …”), define the metric, minimum duration, stop rule and roughly the sample I need, and give three reasons the result could mislead.
Editing and fact-checking (46–50)
The last step before anything goes out. Nothing ships without your sign-off.
Cut this text by 40 % without losing key points, evidence or examples: [text]. Remove filler, repetition and throat-clearing intros. Then list what you cut so I can check it.
[Brand brief]. Check this text against our brief: [text]. Mark passages that don’t sound like us (too salesy, too formal, hype words) and suggest an alternative for each. Rate the overall voice 1–5.
List every checkable claim in this text: numbers, dates, names, quotes, prices, studies: [text]. For each, say what kind of source would back it up and whether the text already cites one. Don’t judge whether it’s true yourself; I’ll check at the source.
Rewrite this for readers skimming on a phone: [text]. Short sentences (20 words max), active verbs, one point per paragraph, subheadings every 3–4 paragraphs, jargon explained on first use. Content and facts stay the same.
Adapt this text for readers in [US/UK]: [text]. Not word for word, but the way a native writer would put it: fitting examples, units, date format and tone. Flag anything that only applies legally or culturally to the original market.
Prompts that fail, and why
Most bad outputs come down to one of five causes. The table shows a typical prompt, the problem and a better version.
| Weak prompt | Why it fails | Better |
|---|---|---|
| “Write an Instagram post about our product.” | No context, goal or audience, so the AI writes the average of all posts. | [Brief] + goal, audience, format, 3 variants |
| “Give me AI marketing statistics.” | Language models invent plausible numbers and sources. | Paste your own sources and have them analysed, or research with citations and then verify |
| “Make this better.” | “Better” isn’t a criterion. The AI changes style, length and sometimes facts. | A concrete goal: shorter, mobile-friendly, voice per brief, facts unchanged |
| “Write an email that sells.” | Leads to pressure, fake scarcity and claims that are legally risky. | An honest block: who it’s for, who it isn’t, dated price, ad disclosure |
| One prompt for everything at once | Long chains of instructions drop requirements along the way. | Split it: outline first, then section by section, then review |
The pattern: the less you say about your brand, customers and goal, the more the AI fills with generic filler. Prompts aren’t magic spells. They’re a brief for someone who writes fast but doesn’t know your business.
Do these work in Claude and Gemini?
Yes. All 50 templates are written without ChatGPT-specific commands. The core advice in the OpenAI, Anthropic and Google guides is the same: clear instruction, context, format, examples. There are small differences:
| Assistant | Where the brief goes | Our experience |
|---|---|---|
| ChatGPT | Project or custom instructions; custom GPTs for recurring jobs | Strong at variants, images and data analysis in one chat |
| Claude | Project with project knowledge | Strong at long-form copy and a consistent voice; we use it for our own writing |
| Gemini | Gems with their own instructions | Handy if your data lives in Google Docs, Sheets and Gmail |
Pricing and free plans for all three: best AI marketing tools.
Tip: run your five most important prompts in two assistants with the same brief. After a week you’ll know which suits your work, without relying on reviews.
Privacy and disclosure
Keep customer data out of consumer chats
On personal ChatGPT plans (Free, Plus, Pro), OpenAI says it may use your content to train its models; you can switch this off under Settings › Data controls (“Improve the model for everyone”). For ChatGPT Business, Enterprise and the API, OpenAI says inputs aren’t used for training by default. Either way: names, email addresses, customer history or internal figures only belong in tools you have a data processing agreement with, which GDPR requires if you handle EU residents’ data.
- Almost none of the 50 templates need personal data. Replace names with roles (“Customer A”).
- Anonymise reviews and survey answers before pasting them in.
- For tasks involving customer data, use a business plan with a contract, or do the task without AI.
Disclosure under the EU AI Act
If you reach people in the EU: since August 2, 2026, realistic AI images, video and audio that could pass as real must be disclosed. Text a human has reviewed and takes editorial responsibility for usually isn’t covered. Details, including what applies to ads, are in our AI marketing guide.
Tools that make prompting easier
Two tools noticeably speed up working with prompts: dictating instead of typing for long briefs, and a writing assistant for the final polish.
| Tool | Best for | More |
|---|---|---|
| Wispr Flow | Dictate long prompts and briefs instead of typing, in any app | Overview · Try →Ad |
| Superhuman (formerly Grammarly) | Check and polish copy before it goes out | Overview |
| ChatGPT, Claude, Gemini | The assistants themselves: pricing and free plans in our tools comparison |
“Try” links are affiliate links (advertising): I earn a commission at no extra cost to you.
Frequently asked questions
What makes a good ChatGPT prompt for marketing?
A good prompt states the role, context, task, format, examples and limits. Context matters most: a brand brief with audience, offer, voice and sample copy. Without it, the AI produces interchangeable average copy.
Do these prompts work with Claude or Gemini?
Yes. The templates use no ChatGPT-specific commands and work in Claude, Gemini and other assistants. Store your brand brief as a project in ChatGPT, as project knowledge in Claude and as a Gem in Gemini.
How do I stop ChatGPT from making up numbers?
Supply the facts yourself and add: “Don’t invent numbers, sources or quotes. If something is missing, write [check].” Then verify every figure at the original source.
Can I put customer data into ChatGPT?
On personal plans, OpenAI says it may use content for training, which you can switch off. Personal data should only go into tools you have a data processing agreement with. Most marketing prompts work fine with anonymised input.
Do I have to label AI-written copy?
Under the EU AI Act, disclosure since August 2, 2026 mainly covers realistic AI images, video and audio. Text a human has reviewed and takes editorial responsibility for usually doesn’t need a label. Platform and advertising rules can add their own requirements.
How long should a prompt be?
As long as it needs to be. With a stored brand brief, two or three sentences for task and format are often enough. Without a brief, a detailed prompt with examples beats any short formula.
Which prompts should I try first?
Start with the brand brief, then prompts for your most frequent task, such as subject lines, 30 post ideas or the monthly report. Use the same prompts for a week and refine them based on the results.
Sources
- OpenAI: Prompt engineering – clear instructions, roles, few-shot examples, context
- Anthropic: Prompt engineering overview – success criteria, clarity, examples, roles, structure
- Google: Prompt design strategies – clear instructions, examples, context, decomposition
- OpenAI Help Center: How your data is used to improve model performance – training on Free/Plus/Pro, opt-out, Business/Enterprise/API
- artificialintelligenceact.eu: Practical guide to Article 50 – disclosure duties from Aug 2, 2026
- HubSpot: 2026 State of Marketing – 42.5 % use AI extensively for content creation







Social media (11–15)
For ideas, hooks and variants. Nothing goes live until you’ve checked it.