Why all AI-built UI looks the same

Amratansh — Thu, 29 Oct 2026 05:14:30 +0000

I'm a frontend developer at a startup, and I've built UI with Claude for a long time. One thing kept bothering me: every screen, platform and component came out looking the same. The cause isn't one model. It's how all of them work, and how we use them.

Here's why, in plain terms:

The typical answer wins. Claude, GPT and Gemini are all trained on huge amounts of existing code. Ask for a landing page and you get a hero with a big title, a lede and two CTAs, because that's the most typical answer. Typical is generic.
Same tools, same look. A large share of public frontend code is built with the same few tools (Tailwind, shadcn, starter templates), and that's very likely what the models learned "a UI" looks like.
Short prompts leave every decision to the defaults. "Make me a dashboard" says nothing about aesthetic, mood, audience or colour, so the model fills every gap with a default.
No references. Designers start with a gallery of references, shortlist a few and iterate. AI starts with a sentence and jumps straight to something that looks finished. You never see other aesthetics.
Blending. Ask for "modern but warm but bold" and you get an average of all three instead of a commitment to one.
Nobody with taste makes the choices. I'd get one prototype from Claude and pick and drop bits of it. I never saw other directions, because Claude doesn't offer them unless you ask, and I didn't know to ask.

So I turned a designer-style workflow into a skill: references first, several distinct directions, then my pick at every stage. The prototypes changed completely. As a non-designer, I was suddenly looking at ideas I'd never have come up with myself.

Here is the Claude plugin, do try it

amratansh12 / directionforge

UI that doesn't look AI-made. A Claude Code plugin that replaces generic AI design (purple gradients, icon grids, Inter everywhere) with a real process: references grouped by style, distinct directions built with impeccable and tasteskill, your Always/Never rules, and your pick at every stage.

directionforge

UI that doesn't look AI-made. A Claude Code plugin that makes AI-generated UI look designed.

See a full run: one brief for a fictional ceramic-studio app, built three ways (a one-line prompt, the full brief, and directionforge), with every direction, iteration and decision.

Ask an AI for a page and you get the usual: purple gradients, glossy blobs, rows of icons with three words each, rounded everything, Inter everywhere, every colour used at once. That's what happens when there's nothing to steer it.

directionforge gives it something to steer by:

Real references. Every direction comes from screenshots you or Claude collect, grouped by visual style.
Several distinct directions. Each one commits fully to its own look, and none of them blend.
Your rules. You write the Always and Never lists, and every build has to follow them.
Your choices. Claude stops at every stage until you pick, and…

View on GitHub

DEV Community: Amratansh

Why all AI-built UI looks the same

amratansh12 / directionforge

UI that doesn't look AI-made. A Claude Code plugin that replaces generic AI design (purple gradients, icon grids, Inter everywhere) with a real process: references grouped by style, distinct directions built with impeccable and tasteskill, your Always/Never rules, and your pick at every stage.

directionforge