library

image

effects

mcp

ai builders · 5 min read

Claude prompts for web design: what makes one work

A Claude prompt for web design works when it is grounded in a real design. Without one, Claude builds from the rules it already has, and those rules are the same for everyone who asks. That is the whole difference between a prompt that gives you a site and a prompt that gives you the site millions of other people got.

Everything else people put in prompts matters far less than this.

Why a good-sounding prompt still gives AI slop

If you start designing straight away inside Claude or Codex, the tool builds from rules it has inside it. Those rules are shared by millions of people making websites with the same tool, and in ChatGPT's case by hundreds of millions of users. So your site comes out looking like theirs, and people spot it as AI-made at a glance.

Describing what you want in more detail doesn't change where the design comes from. The question to answer first is what the prompt is grounded in. We wrote more about the sameness in why every AI website looks the same.

Four ways to ground a prompt in a real design

  1. Send a link to a site you like. Claude will build something close to it. The trouble is whose site it was: if someone finds you copied their design without permission, that is damage to your name at the least.
  2. Send a set of screenshots. This sounds safer and usually fails, for a reason covered below.
  3. Use a prompt written from a finished design. A prompt library gives you designs that were made to be built from, with a license that says you may.
  4. Design the system yourself. If you are a designer, make the design system in Figma, connect Claude to Figma's MCP server, and build from that. It takes the most work and gives a result that is yours.

The third and fourth both work. Which one fits depends on whether you want to design or to build.

Why a pile of references doesn't work

The problem with screenshots is combining them. Give Claude one site and it copies that site. Give it a list of references and ask for something built from all of them, and it has nothing single to follow, so it goes back to its own defaults. The result is visibly made by AI again.

This is why tools and skills that work by gathering lots of references still tend to produce that look. More inspiration doesn't fix it. One finished design does.

What's inside a prompt that rebuilds a design

A prompt from our library is long and exact, because its job is to give every person the same result. The one for DJ Zak, a free website template, is about 7,200 words. Its sections, with how much of the prompt each one takes:

  • Goal: what is being built and how faithful the rebuild should be. About 270 words.
  • Stack and setup: the tools and how the project is laid out. About 450 words.
  • Design tokens: every colour, size, radius and timing as a named value, in a table of 26 rows. About 1,000 words.
  • Assets: the images and files, with where each goes.
  • Layout, section by section: the page described part by part, in seven parts. About 2,800 words, the largest share.
  • Motion: ten separate movements, each with what triggers it and how it behaves. About 1,800 words.
  • Build order, Done when and Make it yours: the sequence to build in, how to tell it's finished, and where your own content goes.

Notice what takes the space. Layout and motion are nearly two thirds of the prompt. A one-paragraph prompt leaves all of that for Claude to decide, and it decides the way it always does.

The full prompt is on the template's page with a free account.

What we stopped putting in prompts

"Act as an expert designer." It does nothing useful. It makes the AI more confident in bad decisions.

Asking for a JSON structure. This sort of worked in 2024. It doesn't make sense now.

Speak naturally for everything after the build

The library prompt is exact so the first build matches the design. Everything after that is different. Once the site exists inside Claude, the fixes go in as plain speech, often dictated with voice-to-text instead of typed. The more naturally you say it, the better the result tends to be.

So there are two kinds of prompt in one project: a long exact one that carries the design, and short spoken ones that adjust it.

How we check a prompt before it goes live

Every template is rebuilt from its prompt before it is published. We run the prompt through Claude Sonnet 5.5 at medium effort first, then through ChatGPT, to confirm it works in more than one place. A prompt that only works in the tool it was written in isn't finished.

The library has 209 unique templates today, and a number of them come in several themes. On a monthly or yearly plan the license covers commercial and client work.

FAQ

What is the best Claude prompt for a website?

One written from a finished design. It names the design's values, describes the page part by part and spells out the motion. A short prompt that only describes a style leaves those decisions to Claude's defaults.

Should I tell Claude to act as an expert designer?

No. In our experience it changes nothing for the better and makes the AI more confident in weak choices. Give it a real design to follow instead.

Do screenshots work as a reference for Claude?

One screenshot of one site gets copied. Several from different sites get averaged back into Claude's usual look, because there is no single design to follow.

Can I use a library prompt for client work?

On uberdesigner, yes, with a monthly or yearly plan. A free account covers personal use.