StyleSeed Forces AI-Generated UIs Through a Score Gate Before Shipping

StyleSeed is an open-source design ruleset that Claude Code, Cursor, and Codex read before generating UI, but its creator found that publishing 74 rules alone wasn't enough — agents would read them and still ship generic-looking first drafts. The fix, introduced in v2.7, replaces the rules-as-documentation approach with an enforced loop: the /ss-build command first locks core design decisions (colors, fonts, corner radii, motion) into a project file, then builds, then runs /ss-score to grade the result across 8 categories on a 0-100 scale. Anything below 80 gets auto-corrected and rescored before the user ever sees it.

Version 2.10 added a visual gate, /ss-verify, which renders the actual screen, takes a screenshot, and has the agent inspect the image directly — because some of the worst AI-generated tells (a broken web font falling back to Times New Roman, dead whitespace, two competing accent colors) only show up in pixels, not in code. When the author ran the gate against his own pre-gate landing page, it scored 58/100, flagging generic icon-chip cards, mockup screens with no focal point, and inconsistent color usage. Mechanical fixes brought it to 86.

A Hacker News commenter later caught accessibility failures the tool itself should have flagged — the score had drifted to 96 after unverified edits. It was fixed to 100 within two hours. The takeaway, repeated twice in the project's own history: a quality gate isn't a one-time check but a loop that must be rerun after every change, even by the people who built it. StyleSeed is MIT-licensed and installable via a single npx command for Claude Code, Cursor, and Codex.