We run our own marketing on AI every day: briefs, competitor profiles, copy critiques, campaign analysis. We have watched, up close, where the value leaks out. The frontier models are extraordinary and getting better every quarter; the weak link is the prompt.
A vague request such as "write me some ad copy for our new feature" returns something bland and slightly off-brand, and the marketer concludes the tool is overhyped. The model was asked to guess. It guessed who the audience is, what the feature does, what the brand sounds like, and what a good answer looks like. Given no information, it produced the statistical average of all marketing copy on the internet, which is the definition of bland.
This guide covers how to structure a prompt, with real before-and-after examples from marketing work, and then how to stop writing prompts by hand and turn the good ones into reusable assets.
The prompt as the weak link
One mental model changes how you write prompts: the model knows language and nothing about your business. When you give it a vague instruction, it fills the gaps with the most probable content it has seen, and the most probable content is the generic content. Everything that makes your answer specific and yours has to come from you, inside the prompt.
A good prompt is a well-briefed request. If you would not hand that instruction to a sharp freelancer on their first day and expect a usable draft, the model will not manage it either. The freelancer would ask you questions. The model guesses and moves on, so the burden sits with you to answer those questions before they are asked.
How to structure a prompt
Across the marketing prompts we run in production, the ones that consistently return usable work carry the same seven things. You will not always need all seven, and the order can flex, but this is the checklist worth running in your head before you hit enter.
Role. Tell the model what role to take. "You are a senior B2B copywriter who has written for analytics platforms" pulls the answer toward a different region of the model's knowledge than no framing at all. It is the cheapest single upgrade to any prompt.
Context. Say who the output is for, what the thing does, and what success looks like. This is where most of the specificity lives, and where most prompts are thinnest.
Task. Start with a verb and name the deliverable. "Write three headline options" is a task. "Help me with headlines" is a conversation.
Examples. One example of a headline you love, or a competitor line you want to beat, teaches the model more about your taste than a paragraph of adjectives.
Constraints. Set the boundaries: word count, reading level, things to avoid, claims you cannot make. Constraints stop the model over-writing and wandering off-brand.
Format. Ask for a table, a bulleted list, three labelled options, or a short email. If you have a downstream use for the output, describe its shape.
Tone. Name the tone you want, such as confident, plain, warm, or technical, and if you can, include an example written in that voice.
Before and after: three examples
The difference shows up on three jobs marketing teams do every week.
Writing a brief
Before:
Write a brief for a blog post about our data platform.
After:
You are a content strategist for a B2B analytics company. We sell a platform that helps CPG brands find where category growth lives. Our reader is a marketing insights lead who is sceptical of AI hype and time-poor. Write a content brief for a 1,200-word blog post on the target keyword "demand space mapping". Include: a working title, the reader's core question, the three sections the post must cover, two proprietary points of view we can add that a competitor could not, and the internal or external link the post should end on. Keep it to one page. Plain, practitioner tone - we write peer to peer, not as gurus.
The first version returns a brief for a blog post that could belong to any company on earth. The second returns something you can hand to a writer.
Critiquing copy
Before:
Is this landing page copy any good? [paste]
After:
You are a conversion copywriter reviewing a landing page hero. Our audience is CMOs at mid-market CPG brands. The one action we want is a demo booking. Read the copy below and give me: the single strongest line and why it works, the three weakest lines with a specific rewrite for each, and one thing a sceptical CMO would not believe. Do not rewrite the whole thing. Be blunt - I would rather hear the problem now than after we ship it. [paste]
"Is this any good" gets you a polite yes and a list of platitudes. The briefed version gets you an edit you can act on this afternoon.
Profiling a competitor
Before:
Tell me about [competitor].
After:
You are a competitive analyst. I am going to give you a competitor's homepage and pricing page. Profile them for a positioning decision, in a table with these columns: who they say they are for, the core promise in their own words, their proof points, their pricing model, and the one gap in their positioning we could own. Flag anything you are inferring rather than reading directly from the pages, so I know what to verify. [paste pages]
The vague version returns whatever the model half-remembers about the brand, which may be wrong and is impossible to check. The briefed version returns a structured read of primary source material, with its guesses labelled.
None of the "after" prompts are clever. They are properly briefed, they take about ninety seconds longer to write, and those ninety seconds are the entire skill.
The scaling limits of hand-written prompts
Once you can write the "after" prompts, you have a new problem: you are writing them from scratch every time. That ninety seconds, times every brief and critique and profile, times every person on the team, adds up to a lot of reinvented wheels. Worse, the quality is inconsistent. Your best prompt-writer produces sharp output and the new starter produces mush, because the knowledge lives in one person's head and their scroll-back history.
You want the good prompt to become the default, so nobody has to remember it or rebuild it. That is the step from prompt to reusable skill.
From prompt to reusable skill
The progression has three steps, and most of the value is in getting past the first.
Prompt. A one-time, typed instruction, fine for a genuine one-off and a poor fit for a repeatable job.
Reusable skill. Take your best "after" prompt for a recurring task (the brief writer, the copy critic, the competitor profiler) and save it as a structured playbook. The role, the constraints, and the output format are fixed. The only things that change run to run are the inputs: this feature, this page, this competitor. Bundle in the reference material the task always needs, like your tone-of-voice notes, your positioning one-pager, or your list of claims legal has approved. Anyone on the team then runs the same high-quality prompt by naming the task, and the new starter gets the same output as your best operator.
Quality gate. This last step separates teams playing with AI from teams running on it. A reusable skill still produces a first draft, and a first draft still needs a check. The answer is a second pass whose only job is to catch what the first pass gets wrong: the AI-tells and stock phrasing, the claim nobody has verified, the line that drifts off-brand. We run our own copy through exactly this kind of gate before anything is published, and it is the reason AI-assisted output can go out under our name without embarrassing us. The generation step gets you to a draft fast; the gate makes the draft safe to ship.
The shift there is from thinking about prompts as things you type to thinking about them as assets you build once and reuse. A team with fifteen well-built skills and a quality gate operates at a completely different speed from a team where everyone is freestyling into a chat box, and the gap compounds every week, because the skills get sharper each time you refine them and the freestyling starts cold every morning.
Starting points
You do not need a platform or a project to begin. Pick the one marketing task you do most often - for a lot of teams that is the brief - and spend an afternoon writing the best possible prompt for it using the structure above. Test it on three real jobs. Refine it. Then save it somewhere the whole team can reach, and agree on one simple quality check before anything it produces goes out.
That single loop, repeated across your five most common tasks, will do more for your team's AI output than any amount of chasing the newest model.
If you want a running start, we have packaged the skills we use on our own marketing (the brief writer, the copy critic, the competitor profiler and more) into a free Marketing Skills Pack you can adapt to your own brand.