Turning Brand Guidelines into an AI Image Brief: The Visual Profile Your Generator Can Actually Follow
Every marketing team that has tried to make AI images “on brand” has done the same experiment. Open the brand guidelines PDF, copy the relevant page into the prompt, generate, and get an image that is somehow both compliant and wrong. The blue is right. The image is still something anyone could have made.
The experiment fails for a reason that is obvious once said. Brand guidelines were written for a designer, and a designer fills in everything the guideline leaves out with judgment. A model fills it in with the average of the internet. As one guide puts it, guidelines are “written for humans, not for machines”: the model interprets “clean, tech-forward, friendly” through its training data, not your brand context. Superside’s 2026 diagnosis of why AI creative still feels off-brand lands in the same place, and the practitioners who have solved it converge on one instruction: use references, not guides.
This post is the translation. It takes the five things in a guideline that do transfer to a generator, shows how to write each in terms a model can follow, gives you the one-page template, and walks through drafting it from your own website in an hour. It is the companion to Visual Brand Consistency at Scale, which covers the audit and the workflow; this one covers the document.
Why adjectives do not transfer
A guideline says “modern and approachable.” Ask a model for a modern, approachable image of a software team and you get the modern, approachable image it has seen ten million times: a bright open-plan office, three diverse people laughing at a laptop, a plant. That is not a failure of the model. It is the model doing exactly what “modern and approachable” means on average.
The same is true of hex codes. #2D5BFF tells a model the color exists; it does not say the color is 60% of every image, appears as a background not an accent, and never sits next to orange. And it is true of the logo rules, clear space, and typography pages, which govern things the generator does not produce at all.
What a generator needs is not a description of your brand’s personality. It is a description of your brand’s images: what kind they are, what is in them, what is never in them, what they look like. Five fields cover it.
The five fields that transfer
1. Palette with roles and proportions
Not a list of hex codes; a sentence about how they are used. “Deep blue is the dominant background or field in most images (roughly 60%). Off-white paper is the secondary (30%). Live green is a small accent (10%), used for one element, never for large areas. No other saturated colors.” A model can follow that. It cannot follow a swatch page.
Include what colors mean: “green marks the thing that is working or live” is a rule that survives across hundreds of images.
2. Imagery kind, per channel
The single highest-leverage decision, and the one guidelines almost never make: is this brand’s imagery photographic or illustrated, and does that change by channel? Most B2B brands want illustrated for abstract concepts on LinkedIn and email and photographic only where something real exists to show. Most local businesses want photographic, and real, for anything involving the team or the work. Write it down per channel:
Social and email: illustrated, flat, limited palette, no gradients. Landing pages: photographic only when we have the real photo; otherwise illustrated. Never photorealistic renders of people or our team.
Then say what “illustrated” means for you, in visual terms: line weight, flatness, texture, how abstract ideas are drawn. One sentence. The reference images do the rest.
3. Subject rules
What appears in your images, and how. Products: shown from a reference photo, never re-drawn from description. People: real or clearly illustrated, never photoreal strangers standing in for your team. Settings: the environments your customers actually work in, named. Motifs: the one or two recurring elements (a shape, a device, a texture) that make an image recognizably yours. Three to six lines.
4. The never-list
This field does most of the work, and it is the one nobody writes. Off-brand AI output is rarely a wrong color; it is a default the model reached for because nothing told it not to. Name yours. A typical B2B never-list:
No glossy gradients or lens flares. No sparkle or four-point-star icons. No robots, brains, or circuit-board “AI” imagery. No glassmorphism or floating UI panels. No handshakes, no pointing at whiteboards, no diverse-team-laughing-at-laptop. No text in the image. No purple.
Every line removes a category of image you would have rejected anyway. Most tools have a place for this (a negative prompt, an exclusions field, or simply the end of the instruction block), and it belongs on every generation.
5. Three to five reference images
Everything above tells the model what not to do and roughly what to do. The references show it. One subject image (your product or mascot on a plain background), one approved photo-style sample, one approved illustrated-style sample, one palette swatch in your proportions, and optionally one layout sample from your standard social format. Approved outputs only, never aspirational images from another brand. The reference-images guide covers how to choose them and why GPT Image 2.5 made them the main event.
What is deliberately not in the brief. Typography, logo lockups, clear space, tone of voice, mission. Those matter and they live elsewhere: the logo is composited after generation (never ask the model to draw it), headlines stay as live text (not pixels), and voice belongs in the brand voice document, which is this brief’s twin for copy.
The template
Copy it, fill it in, keep it under one page. If it does not fit on a page, nobody will paste it.
VISUAL PROFILE: [Brand] v1, [date]
PALETTE
Dominant (~60%): [color, how used]
Secondary (~30%): [color, how used]
Accent (~10%): [color, used for one element only, means ___]
Never: [colors that must not appear]
IMAGERY KIND
Social / email: [photographic | illustrated] - [one sentence on what that looks like]
Landing pages / web: [photographic when real, else illustrated | ...]
Print / one-offs: [...]
SUBJECTS
Products: [from reference only, never re-drawn]
People: [real photos only | illustrated only | never]
Settings: [the two or three environments our customers actually work in]
Recurring motif: [one shape, object, or texture that is ours]
NEVER
- [default 1, e.g. glossy gradients, lens flare]
- [default 2, e.g. sparkle / star icons, robots, brains]
- [default 3, e.g. glassmorphism, floating UI panels]
- [default 4, e.g. stock clichés: handshakes, whiteboards, laughing at laptops]
- No text in the image. No logo in the image (composited after).
REFERENCES (attach to every generation)
1. Subject: [file] 2. Photo style: [file] 3. Illustrated style: [file]
4. Palette swatch: [file] 5. Layout (optional): [file]Version it. When the never-list grows (it will, every time an image gets rejected for a new reason), bump the version and note the date. A profile that changes is a profile someone is using.
Draft it from your website in an hour
You do not need the guidelines PDF to start. You need your own site and your last twenty published images.
- Screenshot the homepage and two key pages. Full page. This is your brand as it actually exists, which is more useful than the brand as the PDF imagined it.
- Pull the palette from the screenshots. Note which color dominates the area, not which is “primary” in the guideline. Estimate proportions. Note the accent and what it is used for.
- Decide the imagery kind by looking, not by preference. Are the images on the site photos or illustrations? Are the photos real (your team, your product) or stock? If stock, that is the first thing the profile will change.
- List the subjects. What actually appears: products, screens, environments, people. Which environments would your customer recognize?
- Write the never-list from your rejects. Pull the last twenty images you published or nearly published, grid them (the 20-image audit), and write down what the ones you dislike have in common. Those are your defaults to name.
- Pick the references. The three images in the grid you would be proudest to repeat. Make the swatch. Save all of them in one folder with the profile.
Then run five generations with the profile and references attached. Grid them. Adjust the never-list. That is version 1, and it is already better than the PDF at the one job the PDF could not do.
The profile has to show up on its own
A profile in a doc works exactly as long as someone pastes it, which is until the first busy week. The durable version is a profile stored with the brand and attached to every generation automatically: palette read from your brand settings, imagery kind chosen per channel, the never-list folded into the model’s exclusions, the reference board rotated in, and a first draft of the whole thing proposed from your own website screenshots so you edit rather than write.
That is what Marqeable’s brand visual profile is. It is the document above, kept live, riding on every image generated inside every email, post, and page, with each image recording which version of the profile and which references produced it. You review it once, adjust it when a reject teaches you something, and never paste it. We are in private beta with a small early cohort: get early access if you want the first draft proposed from your site instead of written from scratch.
Frequently asked questions
Can I just upload the brand guidelines PDF to the model?
It helps less than expected. The model will extract the hex codes and adjectives and ignore the layout and logo rules it cannot use, and the adjectives will be interpreted on average. Extract the five fields yourself; it takes an hour and it is the hour that matters.
How long should the brief be?
One page. The never-list is the section allowed to grow. If the imagery-kind section is more than three lines per channel, you are describing style in words that the reference images should be showing instead.
Does every generation really need all of this?
Yes, and that is the argument for a tool that attaches it for you. The profile is short precisely so it can go on every single prompt without anyone deciding whether this one is important enough.
What if my brand has no visual identity yet?
Then this exercise creates one, cheaply. Pick the imagery kind, pick three colors with roles, write a never-list of the clichés in your industry, generate ten images, keep the three best as references. That is a visual identity a designer can later refine, and it is consistent from day one.
The bottom line
A brand guideline tells a designer who you are. A generator needs to be told what your images are: which colors and in what proportion, photographic or illustrated and per which channel, what appears and what never does, and three approved images that show all of it at once. Five fields, one page, drafted from your own website in an hour, versioned every time a reject teaches you something. Then make it show up on every generation without a person remembering, because the profile that lives in a doc is the PDF all over again.
Marqeable runs your campaigns, answers every visitor, text, and email in seconds, and turns them into booked jobs and meetings - even at 9pm on a Saturday. We’re in private beta with a small early cohort. Get early access
