Taste-Skill is a collection of portable agent skills for improving AI-generated frontend design, plus separate image-generation skills for producing reference boards before implementation.
This repository packages multiple skills intended for AI agents and coding tools. Its main focus is frontend taste: better layout, typography, motion, spacing, and less generic-looking UI. It also includes image-generation skills for website, mobile, and brand-kit reference boards, meant to be used with image tools and then handed off to coding agents for implementation.
The repo targets a common problem in AI-assisted UI work: generated interfaces can look boring, generic, or unfinished. It is also aimed at workflows where teams want a clearer path from brief to visual reference to code, especially when using agentic tools that support reusable skills.
Conceptually, each skill is a reusable instruction set installed through the `npx skills add` workflow or copied directly into a project or chat. The README says the default frontend skill reads a brief, infers a design language, and adjusts three high-level dials: variance, motion, and density; it also includes guidance such as a stricter anti-slop stance, a redesign-audit path for existing projects, and image-first workflows where references are generated before coding. The image-generation skills do not produce code themselves; they are meant to create visual references that another agent can then implement.
It is attracting attention because it sits at the intersection of AI coding, design quality, and agent workflows, which are all active topics in current developer tooling. The repository also appears to be evolving quickly, with the default frontend skill marked as a v2 experimental rewrite and a large set of specialized skills for different UI directions and workflows.
The README points to comparable approaches rather than direct competing projects: Vercel Labs' agent-skills installation flow, ChatGPT Images or similar image generators for reference creation, and coding agents such as Codex, Cursor, or Claude Code for implementation. Within this repo itself, the alternatives are the different skills you can choose based on the task, such as the v1 frontend skill, the GPT/Codex-focused variant, redesign, minimalist, brutalist, soft, output-enforcement, image-to-code, and image-generation skills.
AI-explained · grounded in each repo's README