AI Design Slop: Why AI-Generated UI Looks Generic — and the Fix

You can spot it in a second now. The purple-to-cyan gradient. Glassmorphism with a neon glow. Six identical cards in a row, each with an icon, a heading, and two lines of text. A bounce on every hover. It has a name: AI design slop — the generic, interchangeable UI that coding agents produce by default.

The agents aren't bad at writing code. They're bad at knowing what good looks like — and at checking their own work.

Why AI-generated UI looks like slop

Two reasons, and the second one is the one nobody fixes.

1. No taste. A model will happily generate a hero section, but "looks designed" isn't in its objective. Absent direction, it reaches for the most statistically common pattern — which is exactly the gradient-and-glass aesthetic that floods its training data. Polish without judgment.

2. No loop. This is the real gap. AI design is one-shot: the agent generates, you eyeball it, you fix the same problems by hand, and next session it makes them again. Even tools that critique UI stop at a one-shot report — a list of findings, no fix, no re-check. Nothing closes the loop.

The result isn't just ugly. It's often unusable: missing focus states, contrast that fails WCAG, empty and error states that were never designed. A December 2025 analysis of 470 real pull requests found AI-generated code carried 1.7× more issues and 2.74× more security vulnerabilities than human code. Research presented at CHI 2025 showed AI assistants systematically generate inaccessible markup. The problem is real, measured, and growing as more UI gets generated.

The fix is a closed loop, not more rules

Design rules alone don't fix slop — plenty of linters list what's wrong and stop there. What actually moves output from "AI-made" to "shippable" is a closed loop:

1. Build with guardrails — anti-slop patterns and accessibility enforced at generation time, not bolted on after.
2. Critique the result with real design judgment — anti-slop tells, usability heuristics, an explicit acceptance bar.
3. Fix the highest-impact issue.
4. Re-evaluate — and repeat until the work passes the bar or a budget is hit.

That last step is the whole game. A one-shot generator or a one-shot auditor can't converge. A loop can. It's the difference between "here are 9 problems" and "here's the version with zero problems left."

How UI Craft closes it

UI Craft is a design-quality system for AI coding agents, built around that loop:

• Guardrails — opinionated, craft-level rules across layout, type, color, motion, accessibility, and UX copy, plus an anti-slop test every surface has to pass.
• A 10-item acceptance bar — a concrete, checkable definition of "would a designer retouch this?" The build isn't done until every box is green.
• An iterate-until-converged loop — build → critique → fix the highest-impact issue → re-check, until it passes. The skill grades its own output and keeps going.
• Self-correction — correct it once ("no gradients on the hero") and it records the rule in your project brief, so it never repeats the mistake.
• Portable — it's a skill, not a service. No tool lock-in; if there's nothing to render with, it reasons over the code and tells you honestly what it couldn't verify.

Same prompt, same model — output that clears the bar instead of looking like every other AI demo.

Try it

npx skills add educlopez/ui-craft

Works with Claude Code (native plugin), Codex, Cursor, Gemini, and OpenCode. Full reference at skills.smoothui.dev/docs.

The era of shipping AI slop because "the agent made it" is ending. Hold your UI to a bar — and make the agent clear it.