Teaching AI Agents to Stop Designing Like It’s 2015
Taste-Skill is a collection of portable prompt files that teach coding agents to generate layouts, type, and motion that don’t look like they came from a 2015 Bootstrap tutorial.
Taste Skill is a collection of portable instruction files that teach AI coding agents to generate frontends with actual visual discipline instead of the same three centered feature cards.

What it does
Taste-Skill is a repo of SKILL.md prompt files—essentially glue for your agent’s context window—that steer Codex, Cursor, and Claude Code away from boilerplate UIs. Each file acts as a specialized design director: some enforce layout variance and GSAP motion patterns, others generate reference images for web or mobile comps, and a few audit and redesign existing codebases. There is no runtime library here; just portable instruction sets you drop into a conversation or load through the Vercel Labs skills protocol.
The interesting bit
Instead of one catch-all prompt, the repo splits taste into discrete, tunable dials—DESIGN_VARIANCE, MOTION_INTENSITY, and VISUAL_DENSITY—plus specialist variants like “brutalist” or “soft” UI. There is even a skill whose sole job is stopping the agent from shipping half-finished files with // TODO comments. It treats prompt context like a design system with versioning: v2 is currently experimental, while v1 remains pinned for anyone depending on its exact behavior.
Key highlights
- Framework-agnostic rules target design intent rather than a specific React or Vue API.
- Image-generation skills produce reference boards for ChatGPT Images or Codex image mode, which can then be handed off to coding agents.
- The default skill is a v2 rewrite with a “hard em-dash ban” and canonical GSAP skeletons—suggesting the author has strong opinions on punctuation and animation libraries.
- Includes a dedicated skill whose sole job is stopping the agent from shipping half-finished files with placeholder comments.
- A redesign skill audits existing projects before touching them, rather than assuming greenfield code.
Caveats
- The default skill is explicitly labeled v2 “experimental” and actively iterating, so pinning to v1 may be safer for production workflows.
Verdict Grab this if you are vibe-coding UIs with agents and tired of receiving the same centered-card, gray-background slop. Skip it if you write every component by hand and already trust your own taste more than a prompt file.
Frequently asked
- What is Leonxlnx/taste-skill?
- Taste-Skill is a collection of portable prompt files that teach coding agents to generate layouts, type, and motion that don’t look like they came from a 2015 Bootstrap tutorial.
- Is taste-skill open source?
- Yes — Leonxlnx/taste-skill is open source, released under the MIT license.
- What language is taste-skill written in?
- Leonxlnx/taste-skill is primarily written in JavaScript.
- How popular is taste-skill?
- Leonxlnx/taste-skill has 66.4k stars on GitHub and is currently cooling off.
- Where can I find taste-skill?
- Leonxlnx/taste-skill is on GitHub at https://github.com/Leonxlnx/taste-skill.