What it does

Taste Skill is a collection of SKILL.md files—portable instruction modules—that you install into AI coding agents via npx skills add. Each file tells the agent how to build or design frontends with specific constraints: stronger typography, intentional motion, anti-repetition rules, and layout systems that don’t default to “centered div with a gradient.” There are skills for code generation, image generation, and redesigning existing codebases.

The interesting bit

The default skill (v2, experimental) exposes three tunable dials—DESIGN_VARIANCE, MOTION_INTENSITY, VISUAL_DENSITY—each 1-10, so you can nudge the same agent toward a sparse editorial layout or a dense, animated dashboard without rewriting prompts. There’s even a dedicated skill whose entire job is stopping the model from truncating output with // TODO: implement rest.

Key highlights

• Framework-agnostic: rules target design intent, not React/Vue/Svelte APIs.
• Image-to-code pipeline: generate reference images first, then analyze and implement—useful when you think in layouts, not CSS.
• Specialized variants: brutalist, minimalist, “soft”/premium, redesign-audit, and a stricter GPT/Codex flavor.
• Image generation skills: separate modules for web comps, mobile flows, and brand-kit boards, meant for ChatGPT Images or Codex image mode.
• Google Stitch compatibility: one skill exports optional DESIGN.md format for Stitch workflows.

Caveats

• The default skill is v2 experimental and actively iterating; the author warns it may change before v2.0.0 stable. V1 is preserved if you need to pin behavior.
• The README is clear about what each skill does, but light on concrete before/after comparisons or benchmarked output quality.

Verdict

Worth a look if you’re vibe-coding UIs and tired of manually un-genericizing every agent output. Less useful if you already have a rigid design system and don’t want an agent improvising within it.