A developer who created a rule set to fix the "AI look" in generated UIs found that rules alone are not enough — agents skip the review loop. The solution was to enforce a scoring gate that blocks low-quality output before it reaches the user.
From rules to a forced loop
After publishing a 74-rule open-source design system called StyleSeed, the author noticed that while his own outputs looked polished, others using the same rules still produced generic AI-looking interfaces. The missing piece was not the rules but the process: he had been unconsciously following a lock-build-score-fix loop, while other agents just read the rules and generated a first draft without any review.
To fix this, StyleSeed v2.7 replaced the rule document as the entry point with a mandatory scoring loop. The `/ss-build` command first checks for a design lock file (STYLESEED.md) and refuses to write UI code without it — preventing the drift that occurs when decisions reset on every prompt. After building, `/ss-score` evaluates the output across 8 categories (color discipline, hierarchy, layout, cards, states, motion, consistency, originality) on a 0-100 scale. Any score below 80 triggers automatic fixes and rescoring before the user ever sees the result.
Code passed, pixels failed
Running the gate for months revealed another layer: the worst "AI tells" often don't exist in code but only in rendered pixels. Examples include a hero section where a web font silently failed, rendering in Times New Roman, or a screen with two legitimate accent colors that visually read as two competing highlights. To catch these, v2.10 added a visual gate (`/ss-verify`) that renders the screen, takes a screenshot, and has the agent score the image directly.
Testing the gate on the author's own landing page — built before the gate existed — yielded a humiliating 58/100. The issues were all mechanically fixable: lock an accent color, replace mockups with real product shots, convert icon-chip cards into a numbered flow. The score jumped to 86. Later, a Hacker News comment pointed out accessibility failures on the same site — 96/100 with 19 failures. The lesson: the gate is not a one-time exam but a loop to run after every modification.
Beyond consistency: presets
To address the criticism that consistent UIs can still look generic, v2.8 introduced presets (`/ss-restyle`) — swiss, editorial, technical, warm-dtc, minimal-mono, brutalist-lite. Each preset adjusts corners, type, density, and signature moves as a single coordinate, allowing the entire look to change without breaking consistency. The goal is six coherent, distinct looks that none smell like defaults.
The entire project is open-source (MIT) at github.com/bitjaru/styleseed. A single `npx skills add bitjaru/styleseed` command adds it to Claude Code, Cursor, or Codex.