Top AI Skills Top AI Skills
Comparisons

Anti UI Slop vs Web Design Guidelines: Stopping Generic UI

· 4 min read

Two official agent skills target the same enemy — AI-generated UI that looks like every other AI-generated UI — and they disagree about the fix. Vercel’s web-design-guidelines audits finished code against a live 190-line rulebook and answers with file:line violations. UIZZE’s anti-ui-slop grounds new builds in more than 800,000 real screens, loads one playbook at a time, and closes with a render-and-inspect gate. I read both SKILL.md files and pulled their install curves from skills.sh on 2026-08-25. The numbers sharpen the contrast better than the marketing does.

What Each Skill Actually Does#

web-design-guidelines is a 39-line launcher. It fetches the guidelines file fresh from vercel-labs/web-interface-guidelines before every review, reads the files you name, applies every rule in the fetched document, then reports in file:line format. The rulebook ships separately — 190 lines, 16 rule groups from accessibility down to copy — plus a 14-item anti-pattern list, and the source doc tells the agent to “sacrifice grammar for brevity. High signal-to-noise.” I broke the full rulebook down in the web-design-guidelines intro, so the short version: it’s a compliance sweep you point at existing code.

anti-ui-slop (v1.2.13, MIT, from UIZZE) works the other side of the pipeline. Its opening line is “Stop making UI slop.” The skill reads the product brief, existing components, design tokens, and constraints first — those outrank the skill itself, and it warns against adding “novelty for its own sake.” Then it loads exactly one playbook, chosen from six: new-work for new interfaces, operate for products and dashboards, polish, distill, audit, and a dedicated ios playbook. The optional paid UIZZE MCP adds precisely two tools, find_ui_references and find_ui_materials, and the skill tells the agent never to claim the MCP is connected without an actual host result. It ends with a finish gate: render once when the environment allows, fix observable breakage — clipping, overlap, distorted media, inaccessible controls, inert interactions — and keep the handoff concise.

Same Goal, Opposite Postures#

The sharpest difference is what each treats as the source of truth.

web-design-guidelines treats quality as a compliance problem. Rules get fetched, files get checked against all of them, and the output is complete: findings grouped by file, clickable in VS Code, silent only when a file passes with a single ✓ pass line. Nothing gets ranked, nothing gets capped.

anti-ui-slop treats quality as a judgment problem. Its audit playbook says “review only what can be observed in the implementation or rendered result. Do not invent missing requirements or turn personal taste into a defect.” An audit caps at three material findings, ordered by user impact, and the SKILL.md warns the agent to “apply judgment rather than treating its examples as a checklist.”

One skill is exhaustive. The other is selective on purpose.

The Install Curves Tell Different Stories#

web-design-guidelines: 574,938 installs, rank #37, weekly band between 17,172 and 21,343 across its last eight reports — steady, drifting up slightly.

anti-ui-slop: 481,847 installs, rank #61, and its weekly reports read like a hockey stick: 0, 0, 4,081, 14,588, 86,056, 106,826, 119,426, 121,261. Eight weeks from zero to 121K weekly installs — the steepest growth curve I found among design skills on the board.

The star counts make the contrast stranger. vercel-labs/agent-skills carries 30,412 stars; uizze/uizze has 11. The exploding skill comes from an 11-star repository. Vercel’s name rides the marketplace and its own clout; UIZZE ships through uizze.com plus plugin packs for Claude Code, Cursor, and Codex sitting in the same repo — the wave isn’t coming from GitHub discovery.

So the winner depends on the axis: most total installs is Vercel. Momentum is UIZZE, unambiguously.

Which One to Reach For#

Real talk — these two slot into different moments of the same flow, and I’d run them in sequence.

Sweeping code you already wrote before a release? web-design-guidelines is the tool. It’s exhaustive, it produces a diff-friendly list, and it will flag the missing aria-label on the icon button and the transition: all you forgot to split.

About to generate or redesign a screen? anti-ui-slop shapes the generation itself — it changes what the agent produces before a rulebook gets a chance to complain. Its three-finding audit also reads better in a review discussion than a 60-item file:line dump.

My honest setup: write UI with anti-ui-slop, then sweep the result with web-design-guidelines before shipping. The judgment skill sets the direction; the compliance skill catches the stragglers. Both run free — the UIZZE skill explicitly needs no account, token, or MCP connection for the free tier.

Try the pair against one real component this week. Design a settings screen with anti-ui-slop, audit the output with web-design-guidelines, and count which one catches more. Start with the full web-design-guidelines teardown for the 16 rule groups spelled out, then grab anti-ui-slop from github.com/uizze/uizze or uizze.com.