Key takeaways
- Write prompts like a brief for a capable new hire: goal, context, source material, constraints, output format and an example.
- Paste in your own material instead of asking the model to remember facts. It is far less likely to invent things.
- Treat the first answer as a draft. Critique it, ask the model to ask you questions, and refine in the same conversation.
- When you reuse a prompt every week, turn it into saved instructions, a project, a skill or an agent.
To write AI prompts that work, write them the way you would brief a smart new hire who knows nothing about your company: say what you want and why, give the context and source material, set the constraints, and describe the output you expect. Most weak AI output comes from thin prompts, not weak models. The fix is a habit, and your team can learn it in an afternoon.
This is part 3 of our AI 101 series for marketing teams. If you want to know why a model behaves the way it does before you learn to steer it, start with part 2, how large language models work.
A simple prompt framework for marketing work
A useful prompt answers the same questions a good creative brief answers. You don't need every element every time, but when a result disappoints, the missing piece is usually on this list.
| Element | What to include | Example |
|---|---|---|
| Role and goal | Who the model should act as and what a good result achieves | "Act as a B2B email copywriter. The goal is demo requests from existing customers." |
| Context | Audience, product, stage of the funnel, what has already been tried | "Readers are operations managers who already use our scheduling product." |
| Source material | The facts the output must rest on, pasted in or attached | The product one-pager, the campaign brief, last quarter's report |
| Constraints | Length, tone, words to avoid, legal or brand rules | "Under 150 words. No exclamation points. Don't mention pricing." |
| Output format | The exact shape you want back | "Three subject lines, one preview text line, then the body." |
| Examples | One or two samples of what good looks like | A past email that performed well, with a note on why you liked it |
Order matters less than completeness. Some people put the goal first so they don't forget it; others lead with the source material. Pick one order and use it consistently, so your team's prompts are easy to read and easy to reuse.
A before-and-after prompt example
Here is a hypothetical case. Imagine a mid-sized software company launching a new reporting feature, and a marketer who needs a launch email to existing customers.
Before:
Write a product launch email for our new reporting feature.
The model will produce something. It will be generic, it will guess at features, it may invent a benefit or two, and it will probably sound like every other launch email. None of that is the model's fault. It was given nothing to work with.
After:
You are a B2B email copywriter writing for our existing customers. The goal is to get them to try the new scheduled reports feature in the next two weeks.
Context: our customers are marketing operations managers. Their main complaint, from support tickets, is that they rebuild the same weekly report by hand every Monday.
Source material: the release notes are pasted below. Only use features described in them. If something you want to say isn't supported by the notes, leave it out and tell me.
Constraints: under 150 words, plain and friendly, no exclamation points, no pricing, one call to action that links to the setup guide.
Output: three subject line options, one preview text line, then the email body.
Here is a past email our customers responded well to, for tone: [paste email].
[paste release notes]
The second prompt takes a couple of minutes longer to write. It saves far more than that in editing, and the draft is grounded in real product facts instead of the model's guesses.
Iterate in conversation, and give the model your own material
Treat the first answer as a first draft, the way you would with a junior writer. The conversation is where most of the quality comes from.
- Critique specifically. "Make it better" gives the model nothing. "The second paragraph repeats the first, and the call to action is buried" gives it something to fix.
- Ask the model to critique itself. "Review this draft against the constraints I gave you and list anything that doesn't meet them" catches a surprising amount before you read closely.
- Change one thing at a time when a draft is close, so you can tell which instruction made the difference.
- Start fresh when a thread goes sideways. Long conversations pile up conflicting instructions. Copy the best version into a new chat with a clean prompt.
Just as important: give the model your own material instead of asking it to recall facts. A language model predicts plausible text. Ask it for your product's feature list or last year's conversion rate and it may produce something that sounds right and isn't. Paste in the brief, the report, the transcript or the style guide, and tell it to work only from that. Grounding the model this way is the single most reliable way to reduce made-up details. You still check the output, but you are checking against a source you supplied.
Ask the model to ask you questions first
When you are not sure what the model needs, let it tell you. End the prompt with: "Before you start, ask me up to five questions about anything that would change your answer." The questions often expose gaps in your own thinking, such as an undefined audience or an unstated goal, and the draft that follows is usually much closer to what you wanted.
Copy-ready prompt templates for common marketing tasks
These AI prompt templates follow the framework above. Replace anything in square brackets, and paste your real material where indicated. Keep the ones that work in a shared document so the whole team starts from the same place.
Campaign brief to email draft: You are an email copywriter for [company], writing to [audience]. Using only the campaign brief below, draft one email whose goal is [goal]. Keep it under [number] words, match the tone of the sample email, and include one call to action: [CTA and link]. Give me three subject lines, one preview text line and the body. Flag any claim in your draft that isn't supported by the brief. Brief: [paste]. Sample email: [paste].
Summarize a report for leadership: You are preparing a summary for [role, for example the CMO], who has two minutes to read it. Using only the report below, write five bullets: what happened, why it matters, what changed from last period, the biggest risk, and one recommended decision. Use numbers exactly as they appear in the report and don't calculate new ones. If the report doesn't support a point, say so. Report: [paste].
Create a UTM-tagged link list: Build a table of tracking links for the campaign below. Columns: channel, placement, full URL. Base URL: [URL]. Use utm_source, utm_medium and utm_campaign, plus utm_content where placements need to be told apart. Follow our naming rules exactly: [paste rules, for example lowercase, hyphens instead of spaces, campaign name format]. Channels and placements: [list]. After the table, list any placement where our rules were ambiguous.
Repurpose a webinar into posts: Below is the transcript of our webinar on [topic] for [audience]. Pull out the five most useful, specific ideas a practitioner could act on. For each, write one LinkedIn post of 80 to 120 words in the speaker's own voice, quoting the transcript where you can. Don't add facts, statistics or examples that aren't in the transcript. Transcript: [paste].
Brand voice check: Here are our brand voice guidelines and a draft. Review the draft against the guidelines only. Return a table with three columns: the sentence, the guideline it conflicts with, and a suggested rewrite. Then list any words from our banned list that appear. Don't rewrite sentences that already follow the guidelines. Guidelines: [paste]. Draft: [paste].
Before you hit enter: a prompt checklist
Run through this list for any prompt that matters. It takes less than a minute.
- Did I say what a good result achieves, not just what to produce?
- Did I describe the audience in a sentence?
- Did I paste in the source material, and tell the model to use only that?
- Did I set length, tone and anything that is off limits?
- Did I describe the output format I want back?
- Did I include an example of good work, if I have one?
- Did I remove anything confidential that our AI usage policy says shouldn't go into this tool?
- Do I know how I will check the output before anyone else sees it?
That second-to-last item deserves its own policy. Part 5 of the series covers writing an AI usage policy for marketing teams.
When a prompt should become something reusable
If someone on your team pastes the same long prompt every week, that prompt has become a process, and it should be stored like one. There are a few steps up from a one-off chat, and part 1 of the series explains the differences between AI chat, skills, agents and coding agents in more detail.
- Saved instructions. Most AI tools let you store standing instructions, such as your brand voice or house style, that apply to every conversation. Good for rules that never change.
- A project or workspace. Group the instructions and reference files for one body of work, such as a product line or a campaign, so every chat inside it starts with the right context.
- A skill. Package a repeatable task (the steps, the rules and any reference files) so the model follows the same procedure every time. Our free campaign QA skill is an example: it checks a campaign against a consistent list instead of relying on whoever wrote the prompt that day.
- An agent. When the task needs to pull data, use other tools or run on a schedule without someone typing a prompt, it has outgrown a chat window. See our overview of AI agents for marketing teams.
A reasonable rule: once a prompt has been reused a handful of times and produces consistent results, save it as instructions or a skill. Once it needs to act on its own, consider an agent.
Prompting well is the foundation for everything else your team will do with AI. If you want help turning good prompts into shared templates, skills and working systems, ATL Martech's AI for marketing service starts with the tasks your team already does every week. Next in the series, part 4 covers how to choose AI tools for marketing.
Put this into practice
Need help with ai for marketing?
AI for marketing is the use of machine learning and generative AI to plan, create, personalize, analyze and automate marketing work. ATL Martech identifies the use cases most likely to pay off, implements them inside your existing stack with human review and data safeguards, and measures them like any other investment.