StyleSeed wants AI-built UI to stop looking generic

StyleSeed is pitching itself as the layer that makes agent-generated interfaces look less like they were generated at all. The open-source project, posted to Hacker News via GitHub, bundles 74 design rules, 19 skills, and 7 brand skins for tools including Claude Code, Codex, Cursor, Amp, and Gemini CLI.

The core idea is simple: instead of feeding models more design tokens and hoping for better taste, StyleSeed tries to encode design judgment directly. Its skins are described as inspired-by token sets rather than recreations of companies' design systems, while a separate preset layer handles the broader visual restructuring.

A big part of the pitch is enforcement. StyleSeed says it can run a scored Quality Gate that reviews a UI and fixes common failures before the user sees the result, targeting a score of 80/100 or higher. In the project’s example flow, an agent first locks decisions such as key color, font, radius, and motion into STYLESEED.md, then re-reads that file on every prompt to prevent design drift.

The rules themselves are highly specific. StyleSeed calls for a refined black of #2A2A2A instead of #000, a strict single-accent approach, tabular numbers for changing values, low-opacity shadows of 8% or less, and a ban on visual clichés such as rainbow lists, emoji icons, and the familiar Lucide icon inside a pale chip on every card. It also defines motion tiers, while explicitly banning scroll-jacking.

How StyleSeed is meant to be used

The fastest setup, according to the project, is:

Recommended reading

Photoshop AI tools make editing easier for beginners

• npx skills add bitjaru/styleseed

That installs the optional skills layer for supported agents. From there, users can run commands such as /ss-score, $ss-score, /ss-build, $ss-build, or /ss-setup depending on the tool. If installation is not possible, the markdown rules can still be copied in manually through files such as CLAUDE.md, AGENTS.md, .cursorrules, and DESIGN-LANGUAGE.md.

The repository argues that this distinction matters. Without the installed skills, the “gate” becomes more of an honor-system self-check that models may skip. With the skills in place, StyleSeed says the scoring-and-fixing loop actually runs before output is shown.

What it claims to solve

The project frames itself as different from brand token libraries, “make-it-prettier” prompts, and UI generators such as Claude Design or v0. Those tools may help with scaffolding or styling data, StyleSeed argues, but they do not explicitly target the recurring tells of model-generated design or enforce consistency across sessions.

Its claim is narrower than a full design system and broader than a theme pack: a portable rules engine for agent-driven UI work, released as free software under the MIT license. For developers already using shadcn/ui or agent-based coding workflows, the appeal is straightforward: keep the speed, lose the default-indigo dashboard look.