How to Describe a UI to AI — 6 Steps From «Make It Pretty» to a Working Screen

You ask for "a nice profile screen" and you get a nice profile screen. Tidy, even, useless. Because the moment real data shows up, it falls apart: a name wraps to three lines, there's no avatar, the list is empty, the request failed.
The thing is, the model gave you exactly what you described: a picture under ideal conditions. But an interface isn't a picture. It's a picture plus everything that can go wrong.
NN/g even created a separate deliverable for briefing generative AI — the promptframe: "a design deliverable that documents content goals and requirements for generative-AI prompts based on a wireframe's layout and functionality." The key word there is "requirements." Not "wishes."
Below are six steps that turn a request into those requirements.
Step 1. Start with the user's task, not the screen
Your first line shouldn't describe a screen. It should describe a person and their goal.
Bad: "make a settings page." Good: "the user wants to turn off notifications and leave immediately; visits rarely, usually on a phone."
That's not fluff. From that sentence the model can infer that the main action belongs on the first screen without scrolling, not buried in a third tab.
Step 2. Describe the structure top to bottom
Next, a dry list of blocks in the order they appear. You're effectively drawing a wireframe in words:
1. Header: avatar, name, "Log out" on the right
2. "Notifications" block: 3 toggles with labels
3. "Account" block: email (read-only), change password
4. Bottom: "Delete account", muted, separated from everything
One line per block. The model reads hierarchy from order and nesting, not from adjectives.
Step 3. List the states — this is where the difference lives
This is the step nearly everyone skips, and it's exactly what separates a mockup from a working screen. For every block that holds data, say what to show when the data is missing or something broke:
- empty — "no notifications" plus a note that this is normal;
- loading — a skeleton, not a spinner bouncing in the middle;
- error — the error text plus a "Retry" button;
- a lot — what happens with two hundred rows;
- edges — a 60-character name, a missing avatar, a very long email.
Real users spend a serious share of their time in these states. By default the model won't draw a single one — nobody told it they exist.
Step 4. Replace adjectives with tokens
"Modern," "minimalist," "stylish" are words a model understands as "average internet." They don't narrow the choice, they widen it.
Swap them for concrete values — design tokens:
Colors: bg #141416, text #F2EDE4, accent #D4A373, muted #8A8580
Spacing: only 8 / 16 / 24 / 32
Radius: 12px everywhere
Type: one family, three sizes — 14 / 16 / 24
Fifteen lines, and they move the result further than a paragraph of epithets.
Step 5. Give a reference — and an anti-reference
References beat descriptions: "like Telegram's settings" saves a paragraph. NN/g lists examples as one of the required elements of a promptframe, right alongside objectives and constraints.
The anti-reference is just as useful and oddly rare: "no gradients," "no card inside a card," "no emoji icons." You're crossing out exactly the clichés the model drifts toward by default.
Step 6. One screen at a time — then lock it in
Don't ask for three screens in one request: the model will start rationing attention and simplify all three. Once one is done, restate the good decisions in text ("here's our spacing and states, keep using them") and carry that into the next request. Otherwise every new screen starts from a blank slate.
What you get
The difference between two phrasings of one task:
Make a beautiful modern profile screen for my appThe second prompt is several times longer, and that's fine. You're not describing a picture, you're writing a spec. The underlying logic matches how to write a good prompt, just adapted to interfaces.
Do I need to draw a wireframe if I'm describing it in words anyway?
If the screen has more than five blocks, yes — even on paper. A one-minute sketch exposes structural holes that text hides. NN/g stresses that a promptframe doesn't replace the wireframe, it supports it.
Why does the model ignore some of my requirements?
Usually there are too many in one request and they contradict each other. Split it: structure and states first, style in a separate message. And check the result against your own list, point by point, rather than by eye.
Can I just hand it a screenshot?
Often you should — a model reads layout more accurately from an image. But a screenshot carries no states and no rules; what happens on an empty list or an error still has to be written out.
The AI Architect 2027 course: step by step from computer basics to your own app and first customers. 543 short lessons on your phone.





