Taste-Skill is a JavaScript-based framework designed to improve the visual quality of AI-generated user interfaces by providing portable agent skills that enhance layout, typography, motion, and spacing. The project addresses a specific problem in AI-assisted development: preventing generated interfaces from appearing generic or low-effort. Built primarily by Leonxlnx with contributions from optimization2026 and Blueemi, the repository has grown to 56,421 stars and 3,858 forks as of its most recent tracking update.
The core offering consists of multiple specialized skills that developers can integrate into AI coding agents like Claude Code, Cursor, and Codex. The default skill, design-taste-frontend, recently underwent a substantial rewrite to version 2 (currently experimental), introducing features like brief inference, design-system mapping, canonical GSAP code skeletons, and a redesign-audit protocol. The original v1 remains available for projects requiring its exact behavior. A stricter GPT-optimized variant called gpt-taste provides higher layout variance and aggressive anti-slop enforcement specifically tuned for GPT and Codex models.
Beyond the core design skills, the repository includes specialized variants for different use cases. The image-to-code-skill enables an image-first pipeline where developers generate site references, analyze them, and then implement frontends to match. The redesign-skill audits existing projects before improving layout, spacing, hierarchy, and styling. Visual direction-specific skills include soft-skill for polished, premium interfaces with softer contrast and spring motion; minimalist-skill for editorial product UI resembling Notion or Linear; and brutalist-skill for hard mechanical design with Swiss typography and sharp contrast. An output-skill addresses incomplete model outputs by enforcing full code delivery without placeholder comments.
The repository also provides image-generation skills that produce design references rather than code. These include imagegen-frontend-web for website compositions with strong typography and spacing, imagegen-frontend-mobile for iOS and Android mockups, and brandkit for brand identity boards covering logos, palettes, and type applications. These image skills integrate with ChatGPT Images and Codex image mode, allowing developers to generate visual references that can then be handed to coding agents for implementation.
Installation occurs through the npx skills CLI, which scans the skills folder and allows installation by the install name field within skill frontmatter. The taste-skill v2 introduces three configurable dials: DESIGN_VARIANCE controls layout experimentation from centered to asymmetric, MOTION_INTENSITY adjusts animation depth from hover effects to scroll-based interactions, and VISUAL_DENSITY ranges from spacious layouts to dense dashboards.
The project maintains active engagement with its community, with a median issue and pull request response latency of 20.6 hours across 22 tracked items. Leonxlnx leads development with 21 recorded events, while optimization2026 and Blueemi contribute with 3 and 2 events respectively. The repository connects to related projects including koala73/worldmonitor, microsoft/vscode, and anthropics/claude-code through overlapping contributors. The project is sponsored by Emil Kowalski's animations.dev and the Vercel Open Source Program, and operates under an MIT license. The homepage at tasteskill.dev provides additional documentation and a changelog tracking version updates and feature additions.
