Why AI-built websites look the same, and what changes it
AI-built websites look the same because a coding agent fills every decision you did not make with the most common answer it has seen: the same component kit, the same type, an indigo button, a gradient behind the hero, a row of rounded cards with soft shadows, stock icons. No single choice is wrong. The sameness comes from nobody deciding. What changes it is deciding: a visual language, real type and spacing rules, and an illustration system, written down as constraints the agent has to follow.
Why AI-generated websites look the same
Defaults are the average
A language model writes the likely next thing. Asked for "a landing page" with nothing else to go on, the likely page is the median of every landing page in its training data. Alan West traces the purple-button habit to Tailwind's bg-indigo-500 in Why every AI-built website looks the same: Tailwind's co-founder once apologised, half joking, for making it the default button colour in Tailwind UI. Take the story as a good illustration rather than a study, but the mechanism is sound. Common code in, common code out.
Everyone builds from the same kit
Most agent-built sites start from the same few libraries: Tailwind, shadcn/ui, Lucide icons, Inter or a system font. These are good tools, and that is the point: they are so good that nobody changes them. Sameness from shared free tools predates AI. A 2024 Hacker News thread, Why do so many YC startups have similar branding?, lands on the same cause. Agents just made the kit faster to reach for.
Decoration with no idea behind it
Gradient washes, glowing blobs, a generic 3D render, a flat cartoon person waving at a laptop. These fill space without saying anything about the product. They are the visual version of filler copy, and readers have learned to spot them. People now have a word for it: on Hacker News the threads say "AI slop" and "tells of a slop UI".
No art direction, no owned system
An agent does not know what your product is about unless you tell it, and it does not remember what it decided on the last page unless the decision is written somewhere it reads. So each page is designed from scratch, from the defaults. A brand with an owned visual system (its own type, its own colour logic, its own pictures) looks like itself on page forty. A brand without one looks like the kit.
How to make Claude Code better at frontend design
None of these is a trick. Each one replaces a default with a decision.
- Decide the visual language before any code. Name the subject, its materials and its world, then choose colours and type from that, not from habit. Anthropic's own frontend-design skill works this way: it asks the model for a written design plan (named colours, type roles, layout ideas) and to revise anything that reads like the generic default before it writes code.
- Make real type and spacing decisions. Pick a typeface for a reason, set a type scale and a spacing scale, and put them in your CSS variables or Tailwind config as tokens. Replace the default palette with your own instead of overriding it page by page.
- Write the constraints where the agent reads them. A short rules file in the repo (a
CLAUDE.md, a design notes file, a skill) beats repeating yourself in every prompt. Include the bans: no gradient washes, no identical card grids, no eyebrow labels over every heading. Be specific; "make it look premium" gives the agent nothing to hold. - Use real content. Real screenshots, real copy, real numbers. A page built around what the product actually does has a shape of its own; a page built around placeholders takes the template's shape.
- Spend boldness in one place. One strong element per page, everything else quiet. Most generic pages are loud everywhere at once.
- Get a designer for the system, even briefly. A few days of a good designer's time on type, colour and illustration rules will carry an agent a long way. The agent is good at applying a system. It is not good at inventing one you will still like in a month.
Where illustration fits
Type and colour make a site feel like a brand. Illustration is where it says what the product is about, in a picture. That is also where AI-built sites are weakest, because the usual answers are stock: a flat-people pack, an icon in a gradient circle, a 3D blob.
A good product illustration is a concept before it is a drawing. It picks one object that stands for the idea: a compass for "you're off course", an empty letter tray for "nothing in here yet", a telescope for "look further". It needs no caption. And it has to stay consistent, so the tenth one looks like it came from the same hand as the first. That consistency is the part an agent cannot give you on its own; it needs rules. There is more on that in how to create consistent illustrations with AI.
Other honest routes exist: commission an illustrator for a small set and hand the agent the files; build a strict icon-and-diagram language in SVG; or use typography alone, which is often enough for a statement section.
One concrete example: Linejig
Linejig is our take on the illustration part. It is a library of interactive 3D line illustrations for product moments (empty states, 404s, onboarding, feature explainers, heroes), drawn in three.js. Each figure is a small real object in fine line art that you can turn, hover and click, and one hero part answers the pointer.
The figures follow written rules: one camera, a handful of rounded solids, one line width, a fixed grey ramp, one bright part, motion that does what the real object does. The free Line art skill for Claude Code carries the same rules, so a new figure drawn by your agent matches the library. Before it draws, the skill writes three concepts (an object, its hero part, and why), recommends one and waits for you to pick. The idea gets judged before the drawing does.
It will not suit every brand. It is one look, line-art objects on a plain plate, and it draws objects rather than people, animals or scenes. If your brand is painterly or character-led, an illustrator or an image model is a better fit. And it adds three.js to the page, which matters if you are counting kilobytes; the docs give the weights.
Mariner's compass Free
You're off course; here's your bearing.
Letter tray Free
Nothing in here yet.
Refracting telescope Free
Look further.
Next step
If illustration is the weak spot on your site, browse the library to see the style in motion, or start with the free figures, which embed with no account.
Published 21 November 2026.