Prompts for AI web design that doesn't look generic.
AI designs generic when you give it adjectives. It designs better when you give it decisions: the typeface, what to avoid, what to show and a reference.
"Make it more modern." "Make it look professional." "Cleaner." These are the most common requests and the ones that work worst, because the model turns them into the most average thing it knows. Exactly what you were trying to avoid.
These prompts work in Lovable, v0, Bolt, Cursor, Claude or ChatGPT. Fill in the brackets and paste them one at a time: one change per prompt beats a list of ten.
The brief, before anything else
Paste this when you start a project or before asking for any design change. It hands the model the decisions it would otherwise make on its own.
You're designing the website for [what your business is], for [who your customer is].
Decisions already made:
- Type: [name] for headings and [name] for body. Don't use Inter or the system font.
- Color: [color] background, [color] text and one accent: [hex]. No gradients.
- References: [two or three sites you like]. Take their heading scale and the space between sections, not their colors.
Don't:
- Make sections of three cards with an icon on top.
- Invent testimonials, numbers or customer logos.
- Put small uppercase labels above every heading.
- Animate things on scroll or make cards lift on hover.
- Write copy like "unlock", "supercharge" or "take it to the next level".
Before writing any code, tell me in three lines what page structure you propose and why.For the headline
Rewrite the hero headline. It has to say what the visitor gets and who it's for, in under ten words, with no generic promises. Give me five options that are genuinely different from each other and don't change anything else on the page.Pick one yourself. The model tends to lead with the safest option.
For type and scale
Change the heading typeface to [name] and the body to [name]. The hero headline should be much bigger than it is now (around [size] on desktop), with 0.95 line height and -0.03em letter spacing. Don't use Tailwind's default sizes for headings: define your own.To get rid of the three cards
The [name] section has three identical cards with icons. Replace it with [a list with one sentence per item / one concrete example with a screenshot / a before-and-after comparison]. No decorative icons and no card backgrounds.For anything made up
Go through the whole page and find testimonials, numbers, customer logos, security badges, countdown timers or reviews that I didn't give you. List them and delete them. Wherever that leaves a gap, tell me what real content would need to go there.For motion
Remove scroll-triggered entrance animations and hover effects on elements that aren't clickable. Keep transitions only on buttons and links, 150 to 200 ms, and add a prefers-reduced-motion rule that turns them off.Save the rules in your project
If purple comes back every time you ask for something, it's because the model doesn't remember what you decided. Save the brief where the tool always reads it: an AGENTS.md or CLAUDE.md file at the project root, Cursor's rules, or the project knowledge section in Lovable.
Write it as short rules you can check ("never three icon cards") rather than wishes ("make it feel elegant").
References instead of adjectives
A screenshot of a site you like tells the model more than any adjective. Give it two or three and say what to take from each: one site's heading scale, another's use of space. Ask it to copy one wholesale and you'll get a worse clone back.
After each change, check the result on your phone as well as your laptop. And if you want to know what's left, run the site through the Slop Score.
Frequently asked questions
What prompt stops AI from designing a generic website?
One that gives decisions instead of adjectives: which typeface, which color, which references to look at, and a list of things not to do, like three icon cards or invented testimonials. Ask it to propose the structure before writing code.
Why does the AI keep adding purple even after I tell it not to?
Because it doesn't remember what you decided in earlier messages. Save your rules in a file the tool always reads, like AGENTS.md, Cursor's rules or Lovable's project knowledge.
Do the same prompts work in Lovable, v0 and Cursor?
Yes. Where you store the rules changes, but what works is the same: concrete decisions, a list of don'ts, references and one change at a time.
Should I ask the AI to copy a site I like?
Not the whole thing. Tell it what to take from each reference, like the heading scale or the spacing between sections. Ask for a clone and you'll get a worse copy with nothing of yours in it.