AI search & GEO
Prompt Engineering for Marketers: How to Write Prompts That Work (Tested in 4 AIs)
The 5-part structure that turns a vague request into a clear brief, a test with the same marketing task in four AIs (including what still went wrong) and a template to save your best prompts.
Prompt engineering is the skill of asking an AI for things so it gets them right the first time: who you are, what you want, which facts to use and in what format. It isn't magic or code. It's a good brief.
We checked it with the same marketing task in ChatGPT, Gemini, Claude and Perplexity. With a lazy prompt, all four added emojis, none followed the format and two invented a giveaway. With a 5-part structure, all four did what we asked and none invented a prize.
Here's the structure, the steps, the full test with what still failed, and a prompt template to save your best ones. Of all our AI search and GEO guides, this is the skill that pays off fastest.
On this page
What is prompt engineering?
Prompt engineering means designing and testing instructions so an AI model gives you the answer you need, in the format you need.
Prompt engineering is the practice of writing, testing and refining the instructions you give a large language model like ChatGPT, Gemini or Claude. The goal is an answer that is useful, accurate and repeatable: right today and right next week, no matter who on your team runs it.
What it isn't: a list of secret formulas or a trick to fool the AI. If a prompt only works with magic words, it isn't a good prompt.
Prompt structure: the 5-part prompt framework
A good prompt has five parts: role and context, task, data, format and constraints. You can skip one, as long as you know what you lose.
| Part | What goes in | Example for a bakery |
|---|---|---|
| Role and context | Who's speaking, to whom, where | Social media manager for a neighborhood bakery in Austin; customers 30 to 55 |
| Task | What you want, with a verb | Write 3 Instagram post ideas to launch pumpkin pie preorders |
| Data | What the AI can't know | Preorders open Nov 1, $28 per pie, pickup only |
| Format | How you want it | Table: post type, hook, 60-word caption and call to action |
| Constraints | What you don't want | No emojis, no superlatives, don't invent promotions |
This prompt framework builds on the official guidance. Google's Gemini guide talks about four areas: persona, task, context and format. We split the data out of context, because that's where marketers slip, and we add constraints.
How to write a prompt step by step
To write a prompt that works, fill in the five parts, ask the AI to flag missing info and edit on top of the first answer.
-
Pick one concrete task
«Social ideas» is a wish. «3 posts to launch pie preorders on Instagram» is a task.
-
Write the role and context
Who's speaking, to whom and where. One line is enough.
-
Paste the facts the AI can't know
Prices, dates, area, how to buy. Whatever's missing, it will invent.
-
Ask for the exact format
Table, list or text; word count; columns. That way you can use it without rework.
-
Set the constraints
Tone, banned words and «don't invent facts I didn't give you».
-
Add «if you're missing a fact, ask me»
This is the line that fixes what structure alone can't.
-
Edit on the answer, then save the prompt
Say what to cut and what's missing. When it works, save it to your template.
Role and context: you are [role] for [business] in [city]; the audience is [who]. Task: [verb + deliverable]. Data: [prices, dates, how to buy, anything the AI can't know]. Format: [table, list or text; length]. Constraints: [tone]; don't invent facts, prizes or promotions. If you're missing a fact, ask me before you write.
How to prompt ChatGPT, Gemini, Claude and Perplexity: we tested it
The structure works in all four, but each has a quirk: Perplexity searches the web even when you don't need it.
| Lazy prompt (87 characters) | 5-part prompt (494 characters) | |
|---|---|---|
| Followed the table format | 0 of 4 | 4 of 4 |
| Emojis | 4 of 4 | 0 of 4 |
| Captions within 60 words | No limit | 12 of 12 (32 to 51 words) |
| Invented giveaways | 2 of 4 | 0 of 4 |
What their own guides say, in short:
- ChatGPT: OpenAI recommends separating instructions, examples and context with headings or tags.
- Gemini: in its prompting strategies, Google advises putting context first and the instruction last for its newer models.
- Claude: the Anthropic best practices ask you to be clear and direct and to explain why you want something.
- Perplexity: if you don't need sources, say so. Its docs warn that vague inputs lead to vague searches.
Prompt engineering examples for marketing, plus a prompt template
The best ChatGPT prompts and AI prompts are your own, tested on your data and saved with their results.
Three marketing prompts you can adapt:
- Meta description: «Write 3 meta descriptions of 140 characters for this page [paste text]. Keyword: [keyword], at the start. No promises the page doesn't keep».
- Reply to a 2-star review: «You own [business]. Reply to this review [paste it] in 60 words: thank them, own the specific mistake, say what you're changing and don't argue».
- Meta ad copy: «Write 3 primary texts of 125 characters for [product, price, audience]: one with a stat, one with a question, one with this real testimonial: [testimonial]».
If the output sounds robotic, run it through our method to humanize AI text. And if the prompt will live inside a bot, start with what is a chatbot and what it should never say.
To build a prompt library, a sheet with these columns is enough:
| Name | Goal | AI and model | Date | Prompt | Result | What to change |
|---|---|---|---|---|---|---|
| Cake posts | 3 Instagram ideas | Perplexity Sonar | late October 2026 | (the 5-part one) | Faithful to the data, 32 to 38 words | Add the ordering channel |
Prompt engineering techniques: few shot, chain of thought and meta prompting
Three techniques cover almost everything a marketer needs: give examples, ask it to reason step by step and have the AI write the prompt.
What is few shot prompting?
It's giving the AI examples of what you want before asking. Paste three email subject lines that worked and the fourth comes out in your style. Anthropic recommends 3 to 5 varied examples; Google warns that too many make it copy instead of understand.
What is chain of thought prompting?
It's asking the AI to reason step by step before answering: «first calculate the margin, then compare with competitor prices, then recommend». Use it for math and decisions, not for a caption.
What is meta prompting?
It's using a prompt to create prompts: «ask me whatever you need to write the best prompt for [task]». Great when you don't know where to start, and the official tools in the prompt generator section below build on the same idea.
What is a system prompt?
It's the standing instruction that shapes every reply of a chatbot or custom assistant: who it is, how it talks, what it never does. Same five parts, written once. Same logic as GEO vs SEO: clear, specific information beats templates.
Prompt generator: the official ones and a free DIY version
A prompt generator turns a rough task into a structured prompt. Treat it as a good first draft, never a finished one.
Three official options exist:
- OpenAI: the Generate button in the Playground builds prompts, functions and schemas from a description of your task, using meta-prompts based on OpenAI's own best practices.
- Anthropic: the prompt improver in the Claude Console, open to all Console users, rewrites an existing prompt and lets you tell Claude what to fix.
- Google: the Gemini guide for Workspace suggests asking Gemini to improve your own prompt, then checking that the result still says what you meant.
None of them knows your facts. In our test, all four AIs invented an ordering channel because we hadn't given one, and a generator can't fix that for you. It fills the structure; the prices, dates and channels stay yours.
If you'd rather skip the Console and the Playground, paste this meta prompt into any chatbot:
«I want to [task] for [business]. Ask me up to 5 questions you need to write the best possible prompt, one at a time. Then write the prompt in five parts: role and context, task, data, format and constraints. Put [BRACKETS] wherever a fact is missing.»
A prompt generator earns its keep when it asks you questions. If it only rephrases your sentence, skip it.
Context engineering vs prompt engineering: what changes for marketers
Prompt engineering is how you phrase the request. Context engineering is deciding what information the AI has in front of it every time it answers. Anthropic laid out its view in September 2025: strategies for choosing and maintaining the best information while the model works. The prompt is one piece of that.
| Prompt engineering | Context engineering | |
|---|---|---|
| What it decides | How you ask for the task | What the AI knows when it answers |
| How often | Every request | Once, then review each quarter |
| Example | The 5 parts above | A one-page brand brief |
Our test shows why it matters. With the 5-part prompt, all four AIs still guessed the ordering channel, because nobody had given it to them. A fixed brief would have covered it.
Two of Anthropic's rules carry over to marketing:
- Less is more. The more text you load, the worse the AI recalls what's inside, an effect Anthropic calls context rot. Aim for the smallest set of useful facts. A 300-product catalog won't help you write one post.
- Examples beat rule lists. Anthropic prefers a few varied examples to a long list of exceptions.
The brief for our made-up bakery fits in six lines: who we are, audience, tone, current offers, how to order and what we never say (prizes, fake urgency). Each daily prompt then adds only the task and the day's fact.
Frequently asked questions about prompt engineering
What is a prompt?
A prompt is the instruction you give an AI: a question, a command or a full brief with data. It can be one line («summarize this email») or a structured request with role, facts and format. There's also the system prompt, which sets how a chatbot behaves in every conversation. The clearer and more specific the prompt, the less the AI has to guess, and the less it makes up.
How do you prompt ChatGPT?
Give it the whole brief, then have a conversation. Start with who you are, what you want, your data and the format. If the first answer misses, don't start over: tell it what to cut and what's missing («shorter, no superlatives, use the price»). For big tasks, ask it to ask you questions first. And if a thread gets tangled, open a new one with a better prompt instead of piling on corrections.
Where can I find good ChatGPT prompts?
Lists of «100 ChatGPT prompts» are fine for ideas, but without your data they give generic answers. The best sources are the official guides from OpenAI, Google and Anthropic, plus your own prompt library: every prompt that works, saved with its goal, the AI you used and the result. After a month you'll have twenty prompts tested on your business, worth more than any list.
What are the 5 P's of prompting?
There's no official «5 P's»: different courses and blogs use different lists of five words that start with P. The official guidance is simpler: Google's Gemini guide names four areas (persona, task, context and format), and OpenAI and Anthropic stress clear instructions, context and examples. Use any list that reminds you to include who, what, which facts and in what shape. The letters don't matter; the missing data does.
What are the 7 things of a prompt?
Again, there's no standard, but a complete prompt usually covers seven things: role, audience, task, context or data, format, tone and constraints. You won't need all seven every time. A quick rewrite needs a task and a format; a campaign brief needs all of them. In our test, the part that mattered most for marketing was the one people skip: the facts the AI can't know, like prices and dates.
How to write a good prompt for research?
Ask a precise question, say what sources you trust and how fresh they must be, and ask for links next to every claim. For example: «Find studies from 2024 or later on email open rates for retailers; list 5 with the source, the sample size and the link». Use a tool that searches the web, like Perplexity or ChatGPT with search on, and open the links yourself: AIs still cite pages that don't say what they claim.
What are the best practices for using a system prompt?
Keep it specific enough to steer the AI and loose enough to adapt. Anthropic calls this the right altitude: no brittle if-then rules, no vague «be helpful». State the role, the audience, the tone and what the assistant never does, then add a few varied examples instead of a list of exceptions. Google's Gemini guide adds a useful clause for business bots: tell it to answer only from the context you provide. Test it with ten real customer questions and rewrite whatever breaks.