How to write a "never do this" list for AI images
A never-do list names the things a model tends to add when your brief leaves a gap, and the things your brand, legal team or category rules forbid. Group it by category: subjects, props, colors, moods, styling, composition, text, other brands and claims. Use it to check images. In the prompt itself, describe what you want instead, and keep only a few short "no" lines, such as no text or watermarks.
Why a model needs a never-do list
A model fills every gap in a prompt with something. Leave the surface unnamed and it picks one. Leave the props open and it adds some, often the most familiar ones for the scene: coffee beans around a coffee bag, lemons beside a drink, confetti at a party. Each brand has its own list of clichés to keep out.
The list also covers things that must not appear for other reasons: a competitor’s logo on a prop, a before-and-after the product cannot deliver, a model posed to look underweight. Those are brand, legal and platform problems, and a reviewer can catch them only if they are written down.
The nine categories
| Category | What goes in it | Example lines |
|---|---|---|
| Subjects | People, animals or things that never appear | "No children." "No pets unless the product is for pets." |
| Props | Objects the model likes to add | "No coffee beans, wheat or scattered ingredients that are not in the product." |
| Colors | Colors outside the palette, or ones a competitor owns | "No neon." "No saturated red." |
| Moods | Feelings the brand avoids | "Nothing gloomy. No lonely figures." |
| Styling | Surfaces, fabrics, looks | "No glossy plastic sheen. No marble. No heavy makeup." |
| Composition | Framings the brand avoids | "No tilted horizons. No product dead center in lifestyle shots." |
| Text | Words in the image | "No words, labels or watermarks in the render. Headlines are set later." |
| Other brands | Logos, packaging, recognizable products | "No logos except ours, on our product. No branded phones, sneakers or cars." |
| Claims | What the picture implies | "No before-and-after. Nothing that shows a result the product cannot deliver." |
How to phrase never-do rules in a prompt
A prompt that lists what to avoid still names those things. Google’s prompting guide for its Nano Banana models calls the fix a semantic negative prompt: describe the scene you want, such as an empty, deserted street, instead of writing "no cars". OpenAI’s guide for GPT Image 2.5 keeps a few short exclusions, such as no extra text, logos or watermarks, next to a full description of what it does want. Combine the two. Put positive descriptions and a few hard "no" lines for text and marks in the prompt, and keep the full list in the brief and with whoever reviews.
| Never-do line, for review | What the prompt says instead |
|---|---|
| No clutter | One prop only: a folded linen napkin at the left edge |
| No glossy plastic sheen | Matte ceramic that scatters light softly, no hard highlights |
| No sunset clichés | Flat midday light under a thin, even overcast |
| No ingredients that are not in the product | The bar on a bare slate board, with nothing else on the board |
| No posed smiles | She is reading the label, mouth relaxed |
| No text or watermarks | Keep it as a "no" line: no text, no watermarks, no logos except the one printed on the product |
A never-do list template
NEVER-DO LIST FOR AI IMAGES Brand: ________ Version and date: ________ Owner: ________ Applies to every AI image, before anyone outside the team sees it. SUBJECTS 1. Never show: ________ PROPS 2. Never add: ________ 3. Never add ingredients, parts or accessories that are not in or with the product. COLORS 4. Never use: ________ (outside our palette) MOODS 5. Never feel: ________ STYLING 6. Never: ________ COMPOSITION 7. Never: ________ TEXT 8. No words, letters, numbers, labels or watermarks in the render, except what is printed on the product. OTHER BRANDS 9. No logos, packaging or recognizable products from other companies. CLAIMS 10. Nothing that shows a result, size, quantity or ingredient the product does not deliver. PEOPLE (if they appear) 11. Never pose, light or style a model to look underweight. 12. Never: ________ CATEGORY AND PLATFORM RULES (from legal) 13. ________ WHEN AN IMAGE BREAKS A LINE: reject it and write the line number as the reason.
Examples for food, beauty and fashion
| Category | Lines worth copying | Why |
|---|---|---|
| Food and drink | No ingredients that are not in the recipe. No portion bigger than the pack serves. No steam on food served cold. No glaze or shine the real product does not have. | Pictures make claims too. The US FTC judges an ad by what its words and pictures convey together, and advertisers need proof for implied claims as well as stated ones. |
| Beauty and skincare | No retouched skin in any image that shows a skin result. No filters. No before-and-after unless both are real photos taken the same way. No shade the range does not include. | UK advertising guidance says visual claims must not exaggerate what a cosmetic can do. In 2021 the ASA ruled two tanning ads misleading because filters changed the color of the skin. |
| Fashion | No model posed, lit or styled to look underweight. No fit the garment does not have. No prints, trims or hardware the product lacks. No other brands’ shoes or bags. | The ASA judges how a model is shown, including pose, camera angle, styling and lighting. In February 2025 it upheld complaints about a leggings ad whose angle and styling made the model look unhealthily thin. |
Each category has its own guide: AI food photography, AI photos for beauty and skincare and AI fashion photography. For logos and other brands’ marks, see logos and trademarks in AI images.
Keep the list short enough to use
Aim for a list a reviewer can hold in their head: six to twelve brand lines plus the category rules. Every line should name something visible. "No off-brand energy" fails that test, and "no tilted horizons" passes. Add a line when a fault repeats: if the same reason rejects two images in one round, it goes on the list with the date it was added. The full pre-publish check is the QA checklist for AI images.
Questions people also ask
- Should "no" lines go in the prompt at all?
- A few, for things that are not part of the scene: text, watermarks, extra logos. For the scene itself, describe what you want instead, as Google’s guide recommends. Keep the full list for review, where a person or a checking tool can read each image against it.
- Who owns the never-do list?
- The brand manager, with legal adding the claims and category lines. The creative director should see every change, because each line removes an option from every future shoot.
- How is it different from the brand rules?
- The brand rules describe what the images should look like, and the never-do list names what must never appear. The rules feed the prompts; the list is mostly for review. See how to write brand rules an AI can follow.
Where Overs fits
Overs applies brand rules last so they have the final say, and runs a quality check before any photo renders. When you reject a photo, the reason you give becomes brand memory of what to avoid on later photos.
Free for 40 photos a month. The AI that makes the photos is billed separately, on your own key, with no markup from Overs.
Sources
- Google AI for Developers: Nano Banana image generation, best practices (semantic negative prompts)
- OpenAI: Image prompting, GPT Image 2.5 prompting guide (separate changes from constraints)
- Federal Trade Commission: Advertising FAQs, a guide for small business
- ASA and CAP: Beauty and cosmetics, the use of production techniques (AdviceOnline, August 18, 2025)
- ASA and CAP: Social responsibility, body image (AdviceOnline, February 10, 2026)