跳到正文
DEV Community· Saladi Siddharth·· 12 小時前AI 評分65

Ungeneric 如何修正既有 AI 同質化介面並驗證改進

I built a tool that fixes AI-slop UI and proves it - here is how

AI 導讀

作者推出免費開源工具 Ungeneric,以審查、重建及複查流程修正既有 AI 同質化介面,並量化改進。工具包含 32 項原始碼檢查和 12 項以 headless Chromium 測量的渲染計算樣式檢查,再從 16 種設計系統重組頁面;示例頁面的 slop 分數由 100/100 降至 9/100。整個流程不呼叫模型,同一輸入每次輸出相同,並可離線執行;CLI 亦可用分數上限作為 CI 關卡。

正文

Cover image for I built a tool that fixes AI-slop UI and proves it - here is how

Saladi Siddharth

You know the page before you finish reading it. Purple gradient hero. Glowing cards. Emoji bullets instead of icons. "Supercharge your workflow." They keep happening because a model asked to "make a landing page" samples the safest center of its training distribution - and the center of that distribution is every AI landing page ever screenshotted.

A wave of tools now detects this look. Detection is half the job: after a detector flags your page, you still have to redesign it yourself.

Credit where it is due:

  • uislop - a detector CLI that flags the tells in your HTML.
  • sloplint - a detector skill that tracks even more raw tells.
  • Impeccable - a steering skill (plus detector) that guides the model toward better output.
  • Anthropic's frontend-design - a steering skill for generating new UI from a prompt.

All good tools. None of them rebuilds the page for you. So I built Ungeneric: a free, open-source skill for coding agents that fixes the page and then proves it.

Before: slop score 100/100, purple gradients and emoji icons. After: editorial design system, verified 9/100 on re-audit

The verify loop

Ungeneric treats detection as the first half of a verify loop:

  1. Audit - 32 source checks (gradient washes, glow shadows, stock copy patterns, off-grid spacing, dead anchors and more) plus 12 rendered computed-style checks measured in headless Chromium: real font stacks, contrast, flat type scales, orphaned heading words.
  2. Extract - it pulls the content out of your page: headline, CTAs, stats, features, images.
  3. Rebuild - it recomposes that content into one of 16 real design systems (editorial, brutalist, swiss, night, blueprint, botanical, zine, construct, noir, memphis, letterpress, neon grid, bauhaus, riso print, arcade, manuscript). Your words and links survive; the layout, type, palette and motion do not.
  4. Re-audit - the rebuild goes through the same audit again. If the score did not measurably improve, the CLI exits non-zero and refuses to ship it.

The demo numbers are real: the vibecoded sample page scores 100/100 slop before, and the rebuilt page scores 9/100 after. That remaining 9 is deliberate - it is the page's own claims (its stats, its words) kept on purpose and reported as such, not the tool's aesthetics.

The same vibecoded page rebuilt by all sixteen design systems

Why deterministic matters

There is no model call anywhere in the loop. Same input, same output, every time. That buys you three things:

  • A CI gate - --max-score N exits non-zero above N, so "did this PR make the page more generic" becomes a check, not a vibe.
  • Diffs you can review - rerunning the audit after a change shows exactly what moved.
  • No API cost, fully offline - it runs locally, free forever, no key, no account.

Honest limits

  • Fixed library. The rebuild picks from 16 design systems - predictable and testable, but bounded. For a blank-page brief, a model-steering skill like Impeccable or frontend-design can be more creative. Ungeneric is for the page that already exists and already looks like every other AI page.
  • Heuristic false-positives. Source checks are text heuristics and can flag unusual-but-intentional markup. The score is a conversation starter, not a verdict.
  • Documented disagreements. Where a rule would forbid a whole legitimate aesthetic (neon-on-dark, say), the repo documents the disagreement instead of chasing a zero score. The engine's own output is held to the same standard: all sixteen directions pass every check, with those documented exceptions.

Try it

One install, from any project folder:

npx --yes github:saladi-siddharth/ungeneric

It detects which coding agents you use - Claude Code, Cursor, Codex, Google Antigravity, VS Code (Copilot), Gemini CLI, OpenCode - and installs the skill into each one's skills folder. Reload the agent, then:

use ungeneric on index.html

No agent? The same engine is a plain CLI (node cli.mjs page.html -d swiss -o clean.html), and the repo includes a self-contained local web tool you can open from disk.

Repo: https://github.com/saladi-siddharth/ungeneric - MIT licensed. If you run it on a page, I would genuinely like to know what the audit missed. That is how the check list grows.

來源:DEV Community · dev.to