bitjaru/styleseed

What it solves

StyleSeed addresses the "AI-generated look"—the generic, amateurish UI that coding agents like Claude Code, Cursor, and Codex often produce. It moves beyond providing simple design tokens or palettes by teaching AI agents design judgment and enforcing a set of professional craft rules to ensure products look designed rather than machine-made.

How it works

StyleSeed integrates into AI coding agents via a set of 23 skills and configuration files (CLAUDE.md, .cursorrules, AGENTS.md). It employs a "render $\rightarrow$ score $\rightarrow$ revise" loop where a scored quality gate reviews the generated UI against 74 specific design rules (e.g., banning default indigo accents and emoji icons) and forces the agent to revise any screen that scores below 80/100 before it is presented to the user.

To maintain consistency across sessions, it uses a STYLESEED.md lock file that stores project-specific design decisions (skin, accent color, radius, motion) to prevent visual drift.

Who it’s for

• Developers "vibe coding" SaaS apps who lack a professional design team.
• Users of AI coding agents (Claude Code, Cursor, Codex) who find the default UI outputs generic or amateur.
• Developers using shadcn/ui who want to elevate the output beyond the standard look.
• People creating visual artifacts like social carousels, decks, and reports using AI.

Highlights

• Scored Quality Gate: An automated review process that fixes UI tells before the user sees the result.
• 74 Design Rules: A comprehensive library of professional design judgment rules covering color discipline, spatial rhythm, and information hierarchy.
• Anti-Drift Lock: A STYLESEED.md file that ensures design decisions persist across different screens and sessions.
• Reference Compiler: The /ss-reference skill allows the agent to extract a repeatable design language from screenshots, URLs, or Figma files.
• Multi-Agent Support: Native compatibility with Claude Code, Cursor, Codex, Amp, and Gemini CLI.