How to Prompt AI for Retail Displays: Write a Brief, Not a Prompt

Most people meet AI for display work by typing something like "premium acrylic counter display for a skincare brand, photorealistic" into an image model, looking at what comes back, and deciding that AI cannot do this. I understand the conclusion. I also think the prompt did most of the damage.

I have worked in point-of-purchase for close to four years and I build AI POP Displays, so read this knowing where I stand. This is what I would tell a designer at a manufacturer with an image model open in one tab and a client brief in the other, and it works in whichever tool you already pay for.

Why does a short prompt fail on a display?

An eleven-word prompt has to be completed by something. The model has no failure state, so it never tells you the request was underspecified. It fills the gaps with whatever was statistically ordinary in its training data and hands you a picture at full confidence.

That is fine when the missing parts are decoration. It is expensive here, because on a display the missing parts are the deliverable. Format proportions, how corrugated board or acrylic actually behaves, how many facings really fit, whether the thing reads correctly in an aisle. Those are the parts a client approves and a factory quotes.

So the useful reframe is not how do I write a better prompt. It is what does the model need to know that my customer's email already contains. That is a brief, and you almost certainly have one sitting in your inbox.

There is a second reason this goes wrong, and it is vocabulary. Go looking for prompting advice about displays and you find guides about photographing a product on a clean background, guides about generating a whole store interior, and guides about website pop-ups. Nobody writes about the fixture itself, so the words that carry meaning in this industry never appear in the advice.

What does a display brief have to name?

Seven decisions. I use the same seven whether I am typing into a general model or into our own form, because they are the seven a display designer answers anyway.

1. Sector. Cosmetics reads differently from beverage, which reads differently from pharmacy. Sector carries lighting, finish level, shelf density and the kind of store around it. One word, and it does more work than any adjective you could add.

2. Display format. Counter unit, floor display or FSDU, endcap, shelf tray, glorifier, totem, window, shop-in-shop. These are not synonyms and each carries real proportions. Name the format and then describe those proportions in relative terms, because a general model reads the name as a style cue rather than a geometry.

3. Material. Not "premium" but E-flute corrugated, recycled board, acrylic, brushed aluminium, powder-coated steel, lacquered MDF, oak veneer. Materials behave differently under light and they imply different thicknesses, joints and edges. An adjective gets you a mood. A material gets you a surface.

4. Style. Minimal premium, industrial, sustainable, luxury, editorial, bold graphic. This is the one place a general model needs the least help from you.

5. Retail environment. White studio, soft grey studio, supermarket aisle, pharmacy, beauty boutique, storefront window. Say it explicitly, because "photorealistic" is not a background and the model will choose one for you. Studio when you are reviewing structure, store scene when the client is in the room.

6. Product load. How many SKUs, how many facings across, how many deep, how many tiers. Almost everybody omits this line, and it is the one that makes a render quotable. Leave it out and you get the number of packs that balanced the composition.

7. Framing. Aspect ratio and camera position. A three quarter view at eye height sells a floor unit. A straight elevation is what you want when you are checking whether the structure closes.

For the fuller version of the same discipline, meaning the document you send a manufacturer rather than a model, I wrote that separately in how to write a POP display brief. The seven above are the subset an image model can act on.

This is not a folk theory. Google publishes a prompting guide for its own image models, and its first two rules are to be hyper-specific rather than naming a category, and to explain what the image is for instead of only listing its contents. A display brief is that advice with the industry's nouns dropped in.

What does a vague prompt look like once it is a brief?

Two rewrites, both from real briefs with the names changed.

Before

premium acrylic counter display for a skincare brand, photorealistic, 4k

Six of the seven decisions are open. You will get something acrylic, something countertop, and every other choice made for you.

After

Concept render of a counter display for a premium skincare brand. Sector: cosmetics and beauty, boutique tier. Format: counter unit, roughly twice as tall as it is wide, one sloped tier over a closed base. Material: clear acrylic body with visible edge thickness, brushed aluminium base plate, no fixings on the front face. Style: minimal premium, warm neutral palette. Environment: beauty boutique counter, soft directional light from the left, shallow depth of field. Product load: four SKUs of a 30ml dropper bottle, four facings across, three deep, every front facing forward and legible. Header: brand wordmark centered, header height about a quarter of the total unit height. Framing: 4:5 vertical, three quarter view at counter height.

Same idea, no invented vocabulary, and now nothing important is left for the model to decide.

Before

cardboard floor display for a soda brand, supermarket

After

Concept render of a free-standing floor display unit for a carbonated soft drink brand. Sector: beverage, mainstream grocery. Format: FSDU, three equal shelves plus a header card, footprint proportioned to a quarter pallet, base no narrower than the shelves above it. Material: E-flute corrugated board with a printed laminated finish, fold lines visible at the shelf edges, no metal parts. Style: bold graphic, high contrast. Environment: supermarket aisle end, overhead store lighting, neighbouring shelving visible but out of focus. Product load: 330ml slim cans, eight facings across each shelf, two deep, upright, nothing floating clear of the shelf. Framing: 3:4, eye-level three quarter view.

That is longer than most people are willing to type, which is the honest catch, and I will come back to it.

How should reference images travel?

For anything that has to be reproduced rather than described, words run out. You can describe a bottle. You cannot specify your client's bottle. That is what references are for, and most of the value people lose is in how they attach them rather than in which model they picked. Four rules that hold across tools.

Say what each image is for. An unlabeled pile of images competes with itself. The model has no way to know that image two is a logo to reproduce exactly and image three is a competitor's stand you only want the silhouette from. Write it out. In our own pipeline every reference carries a written role in front of it, one of logo, product photo, campaign artwork, style reference or sketch, because that ambiguity was costing us renders.

Do not overload the stack. References are not free. Google's documentation for its Pro image model puts the ceiling at six object images it will hold to high fidelity, out of fourteen inputs in total. Your logo and four packshots are already five of the six that matter, so a folder of mood shots on top dilutes rather than adds.

Put the structural reference last. These models weight the final image most heavily as the subject to produce. If a customer's sketch or last year's unit is the thing you want followed, it goes at the end of the stack, after the packshots and the logo. We hard-ordered ours that way for the same reason.

Say what to change, not what to avoid. When the reference is a competitor unit or a previous campaign, people write a list of things they do not want, and it works poorly. Google's guide is explicit that describing the absence of something tends to fail and that you should describe the scene you want instead. So rather than "no illuminated crown", write "flat printed header card", and name what you are keeping from the reference in the same breath. The workflow around customer drawings is a subject of its own, and I covered it in sketch to render for POP displays.

How do you prompt a revision without losing the design?

This is where general models cost real hours, and where technique matters most.

Do not rerun the original prompt with an edit bolted on. There is no persistent object underneath the render, so a fresh run resamples everything you did not pin down. You asked for a taller header and the shelf spacing, base width and product load all moved with it.

Feed the approved image back in as a reference, describe the change as a change, and restate the constraints you want frozen. Keep this exact structure, shelf count, material and product load, raise the header by about twenty percent, everything else identical. Google's own guidance points the same way, toward refining conversationally across turns rather than rewriting one long prompt each time. It does not work perfectly, and I would rather say that than sell you a technique.

What will a better prompt not fix?

Prompting is a real skill with a real ceiling, and it is worth knowing where the ceiling sits before you spend a week climbing.

It will not give you exact dimensions. Write proportions, and put the millimeters in the document that goes to your manufacturer.

It will not guarantee an exact brand mark. An approximate logo comes back approximate however firmly you ask, because that is a reproduction problem rather than a wording problem.

It will not make the structure buildable. It will make it look more buildable, which is not the same thing, and the test stays what it always was. Could a manufacturer quote from this without redesigning it. I went through those failure modes in why generic image AI fails at POP displays.

And it will not stop you retyping. That is the real cost of prompt-based display work. The seven decisions are the same seven on every brief, and in a general model you supply them by hand every time, on every variant, on every revision round.

The part I built, and what it is not

Our product is that brief as a form instead of a paragraph. Format comes from a taxonomy that carries proportions, materials come from a list that behaves, environments are options, references travel labeled by role with the sketch last, and all of it gets assembled into a structured prompt you never have to type. The image engine underneath is rented, from Google, and I say so in every post because it matters. You can rent the same one this afternoon. What is not rented is knowing which seven questions decide a display, and that is what close to four years in this industry bought.

Which is also why this post is worth your time if you never sign up. The seven decisions are the job, whatever box you type them into. A designer who carries all of them in their head can hand-prompt to the same place, and plenty do. They just supply it all again on the next brief.

If you would rather answer the questions once and get the render back in under a minute, signup is free with 15 one-time credits and no card required. Pro is $49 a month for 150 concept generations at the founding price, with the $69 list price stated openly. Nothing you upload or render trains AI models, on any plan, because most of this work sits under NDA and that should not be a paid feature.