How to Create Marketing Images With AI: Prompts and Edits
A repeatable three-step routine for turning a plain idea into a usable marketing image, then keeping every image after it consistent with the first.
Generative AI for Marketing: A Workflow, Not a Shortcut
Generative AI does not replace a marketing strategy. It makes a bad one produce output faster, which is worse. Here is the order that actually works.
ChatUp GuidesWork
The most common way generative AI fails in marketing is not bad output. It is good output produced against no strategy — forty social posts, three ad variants, and a landing page, all internally consistent, all aimed at nobody in particular.
Used in the right order, it compresses the parts of marketing that were always slow: research consolidation, positioning drafts, voice documentation, visual concepting, and the long tail of asset production. Here is that order.
Stage 1: research and positioning
Understand the audience with evidence, not adjectives
Prompt to try
Research the market for [product] aimed at [audience]. I want: what problem this category actually solves, how the main players position themselves and in what words, where they overlap, what the recurring complaints are in reviews and forums, and what nobody seems to be saying. Sources for everything, and flag where you're inferring rather than citing.
Find the position you can actually hold
Prompt to try
Based on that research and what we do well — [describe honestly, including weaknesses] — propose four positioning options. For each: the claim, who it wins, who it loses, and what we'd have to be true about our product for it to survive scrutiny. Then argue against the one you think is strongest.
That final instruction — argue against your own recommendation — is how you get past an assistant’s tendency to agree with whatever you seem to want.
Stage 2: brand voice and visual identity
Document the voice so it survives other people
Prompt to try
Here are six pieces of our writing I think sound right: [paste]. Extract the voice into rules a new freelancer could follow: sentence rhythm, vocabulary we use and avoid, how we handle humour, how we talk about competitors, how formal we are with customers. Then write the same paragraph in our voice and in a voice that's subtly wrong, and explain the difference.
The wrong version is the useful artifact. It is much easier to police a voice when you can point at a near-miss.
Visual direction before visual assets
Prompt to try
Give me three visual directions for this brand that are consistent with the positioning we picked. For each: palette with hex values, photography or illustration style, typography character, and what it signals to the buyer. Then tell me which is hardest to execute consistently at volume, because that's the one that'll drift.
Then produce concepts with Image Gen — but treat generated visuals as concepts. For anything appearing as a real product, real people, or real results, use real photography. Composites that misrepresent what a customer will receive are a legal problem before they are a taste problem.
Stage 3: content and campaigns
Copy against a brief, always
Prompt to try
Write the landing page hero for this campaign. Audience: [profile]. Positioning: [claim]. Voice rules: [paste]. The single thing the reader must understand: [X]. Give me five headline options across different angles — problem, outcome, objection, specificity, contrarian — plus a subhead for each. No wordplay, no "unlock", nothing that could headline a different product.
Plan the campaign as a sequence
Prompt to try
Build a six-week launch plan for this. We have two people and about 10 hours a week between us. Give me a weekly sequence with what ships when, what depends on what, and the one thing each week that matters most. Flag where we're likely to fall behind and what to cut first when we do.
Where generative AI is the wrong tool
- Deciding what to sell and to whom. It has no access to your customers, your margins, or your operational reality.
- Anything requiring evidence. Claims, results, testimonials, comparisons. Generate the phrasing, never the substance.
- Regulated categories. Health, finance, legal, and children’s products have rules that a model will breeze straight past.
- Being interesting. Generated copy converges on the median. The line that makes a campaign work is almost always the one a person insisted on.
Measuring whether it worked
Volume of output is not a metric. Track the ones that were metrics before: cost per acquisition, conversion rate, qualified pipeline, retention. If output tripled and none of those moved, the tooling produced more of something that was not working.
Prompt to try
Here are our campaign numbers before and after we changed our process: [paste]. What actually moved, what's within normal variance, and what would I need to measure to distinguish the two? Be conservative — I'd rather hear "not enough data" than a story.
Where this fits in ChatUp
AI Research for stage one, where a stale answer produces a strategy built on last year’s market. The Marketing Expert and Business Planner assistants for positioning and campaign structure. Image Gen for visual concepting. File Assistant for working from your own analytics exports and briefs. AI Goals for the launch sequence, since campaigns fail on execution far more often than on strategy. See also AI tools for marketing and making marketing images.
Frequently asked questions
Is generative AI worth it for a small marketing team?
That is where it helps most — it removes the production bottleneck that keeps two-person teams from testing anything. Large teams gain less because they already had production capacity.
Will customers notice AI-written marketing?
They notice generic marketing, which is correlated but not identical. Copy built from real research and a documented voice, then edited, does not read as generated.
What should I never automate?
Claims, evidence, customer communications about problems, and anything in a regulated category. The rule of thumb: if being wrong would cost you a customer or a fine, a person signs it off.
Where does it save the most time?
Research consolidation, first drafts, variant production for testing, and repurposing one asset across channels. In practice that is most of the calendar.
Order of operations
Strategy, then voice, then production. Reversed — which is the tempting order, because production is the fun part — you get a lot of well-made assets pointing in different directions, and no way to tell which of them worked.
Try it in ChatUp
Turn this guide into a workflow.
Run the prompts above against the model that suits the task, keep the useful context across chats, and pick it back up on any device.
Try for Free