Most AI prompts for marketing produce something that reads fine and means nothing. "Write me five Facebook ad headlines for a skincare brand" gets you five headlines that could belong to any skincare brand on earth. The model isn't broken. The prompt just never told it enough to be specific.
I run paid media across a book of client accounts, and AI does a lot of the first-draft work — ad angles, hooks, copy audits, audience research, reporting summaries. The prompts below are the ones I reach for week to week. None of them are clever. They work because they give the model a role, real context, and hard constraints, then ask for output in a shape I can actually use.
What separates a useful marketing prompt from a generic one
A bad prompt asks for an output. A good prompt sets up the conditions that make the output correct. There are four parts I check for before I run anything:
- Role. Who is writing this — a direct-response copywriter, a media buyer, a data analyst? The role narrows the model's vocabulary and instincts.
- Context. The product, the offer, the audience's awareness level, what's already been tried. Generic in, generic out.
- Constraints. Character limits, banned words, brand voice rules, the framework you want it to follow. Constraints are where specificity comes from.
- Output shape. A table, a numbered list, a JSON brief, copy with a rationale. If you don't specify, you get prose you have to reformat.
Every prompt that follows is built on those four. If you only take one thing from this article, make it the habit of writing all four before you hit enter. It's the difference between a tool and a toy — the same point I made comparing Claude Code vs ChatGPT for marketing work.
Ad angle generation
This is the prompt I run first on any new campaign, before a single word of copy gets written. Angles are strategic positions, not headlines. I want a spread of them so I can test directions, not phrasings.
You are a direct-response strategist who has launched hundreds of
paid social campaigns.
Product: [product + one-line description]
Price: [price] / Offer: [offer]
Audience: [who they are, what they currently do instead]
Awareness level: [unaware / problem-aware / solution-aware / most-aware]
What we've already tried: [angles that worked or flopped]
Generate 6 distinct ad angles. Each must attack the buying decision
from a different place — pain, mechanism, identity, comparison,
objection, status quo. For each angle give me:
- The strategic angle in one sentence
- The emotional driver behind it
- A sample first line (the hook), under 12 words
- Who it best targets
Return as a table. No throat-clearing, no intro paragraph.
The "different place" instruction is what stops the model from handing you six versions of the same idea. Awareness level is the input people skip most often, and it's the one that changes the output the hardest — a most-aware buyer needs a price-and-proof angle, an unaware one needs a problem you name for them.
Hook writing
Once an angle is chosen, hooks are the three seconds that decide whether the ad gets watched at all. I never ask for "good hooks." I ask for hooks built on named patterns, because patterns are testable and "good" isn't.
You are writing opening lines for short-form video ads (TikTok,
Reels, Meta).
Product + angle: [paste the chosen angle from the step above]
Audience: [who]
Write 10 hooks. Use a spread across these patterns: callout,
contrarian, before/after, question, mistake, demonstration.
Label each hook with its pattern. Keep each under 12 words.
Spoken-aloud natural, not written-caption stiff. No emojis.
No exclamation points.
Labeling each hook by pattern does two things: it forces variety, and it gives you a record of which kind of hook won once the ads run. That feedback loop is worth more than any single batch of hooks.
Auditing copy you already have
The fastest AI win in marketing isn't generating new copy — it's pressure-testing copy that's already live. This prompt finds the weak points before you spend money proving they're weak.
You are a conversion copy auditor. Be blunt, not encouraging.
Here is a [landing page / ad / email]:
[paste the copy]
Audience: [who] / Awareness level: [level] / Goal: [the one action]
Audit it against five things and score each 1-5:
1. Hook — does the first line earn the second?
2. Specificity — vague claims vs concrete proof
3. Mechanism — does it explain WHY it works, not just that it does?
4. Single objection — what's the biggest unaddressed doubt?
5. Call to action — clear, single, friction-free?
For each, give the score, the problem, and a rewritten example.
Lead with the lowest score.
"Be blunt, not encouraging" matters more than it looks. Without it the model grades on a curve and tells you your mediocre headline is "strong and compelling." You're not after validation, you're after the one fix that moves the number.
Audience and interest research
Before targeting or messaging, I want the customer's own language. This prompt turns scattered review-and-forum reading into something structured I can feed into copy.
You are a market researcher specializing in customer voice.
Product: [product] / Category: [category]
Here is raw customer language (reviews, Reddit threads, comments):
[paste 15-30 real quotes]
From this, extract:
- The top 5 pains in the customer's exact words
- The top 5 desired outcomes in their words
- The words and phrases they repeat (for ad copy)
- The objections and hesitations they voice
- How they describe the alternatives they've tried
Quote real phrasing wherever you can. Do not invent quotes.
"Do not invent quotes" is a guardrail you need every time you ask a model to summarize source material — without it, it will helpfully fabricate a testimonial that sounds perfect and never existed. The output here feeds directly into the angle and hook prompts above, which is the whole point: these prompts chain.
Turning a week of numbers into a client-readable summary
Reporting is where AI quietly saves the most hours. I pull the raw numbers myself — the model never touches the math — and use it only to translate a table into plain language a client will read.
You are writing a weekly performance note for a client who is smart
but not a media buyer. Plain English, no jargon, no hype.
Here is this week vs last week:
[paste the metrics table: spend, revenue, ROAS, CPA, etc.]
Context: [anything that explains a swing — a new launch, a holiday]
Write 4 short paragraphs: what happened, why, what I'm doing about
it, what to expect next week. Lead with the honest headline, good
or bad. No exclamation points. Don't oversell a flat week.
The rule I never break: the AI summarizes numbers, it does not produce them. Letting a model calculate ROAS or "estimate" a metric is how wrong numbers end up in front of clients. Keep the math in a spreadsheet or a script and let the model do what it's good at — turning it into sentences.
Where prompts alone stop being enough
Run these for a few weeks and you'll hit the same wall I did. You're pasting the same product context into five different prompts. You're re-typing the brand's banned-words list every time. You're copying outputs from one prompt into the next by hand. The prompts work — but you're the integration layer holding them together, and that's a job.
The fix is to stop treating prompts as things you type and start treating them as workflows that live on disk. In Claude Code that means a skill: a folder that holds the prompt, the brand context, the constraints, and any scripts, and fires when you ask for it by name. The product context gets loaded once. The character limits live in the file. One prompt's output feeds the next automatically. I wrote about the first ones worth building in marketing skills for Claude Code.
The production version of this — the full set of ad, copy, research, and reporting workflows wired together with brand context and APIs — is what I packaged into The Operator (course.operatorstack.app/marketers, $397). It's the same prompts above, turned into 60+ skills you run by name instead of paste by hand.
If you'd rather have it run than build it, Clare Digital does this work across client accounts directly. Either way, start with the four-part habit — role, context, constraints, output shape — because no system rescues a vague prompt.
Q: What makes an AI marketing prompt actually work?
Four things, in order: a clear role for the model, real context about the product and audience, hard constraints (character limits, banned words, a framework to follow), and a specified output shape. Generic prompts skip the context and constraints, which is exactly why they produce generic output.
Q: Should AI calculate my marketing metrics?
No. Use AI to interpret and summarize numbers you've already calculated, not to produce them. Models will confidently estimate or miscalculate figures like ROAS and CPA. Keep the math in a spreadsheet or script and let the model translate the result into plain language.
Q: Are these prompts better in ChatGPT or Claude?
The prompts work in either — the structure matters more than the model. The difference shows up when you want to save and reuse them. Claude Code lets you turn a repeated prompt into a named skill that loads its own context, which is covered in Claude Code vs ChatGPT for marketing work.