Taste Skill is a set of Agent Skills that make coding agents design websites instead of assembling the average landing page. The default skill gives the agent an art direction, real pictures and a way to check its own work in a real browser.
Pages it has built, each from one brief:
The photographs are real, found in open archives and credited by the skill (credits). Watch the launch film.
The default skill runs its scripts with Node 22+ and an installed Chrome, Edge or Chromium. It installs nothing and needs no API keys. Finding pictures needs network access.
Claude Code plugin
Run these inside Claude Code. Every skill is namespaced under taste-skill:, so they do not collide with your other skills.
Start a new Codex session afterwards so the skills are loaded.
From a clone, linked instead of copied
./install.sh symlinks the skills into Claude Code (--codex for Codex, --all-agents for both), so a git pull updates them in place. ./install.sh --list shows every install name.
Already have the skill? Run the install command again. The install name did not change, and the older versions stay available under their own names (see Skills).
How v2 works
The default skill is built from a study of 247 curated sites. A build runs in six steps:
Reads the brief and lists what only this client has.
A script deals three of 20 researched directions and the agent picks one.
Finds the idea: one sentence tying the client's own material to that direction.
Builds with real pictures and footage from open collections, credited.
Checks the page in a real browser at three widths and fails it on slop.
Looks at its own screenshots and fixes the top three problems, up to three rounds.
A build takes roughly 30 to 45 minutes and a few hundred thousand tokens. Details are in the changelog and the study.
Skills
Each skill does one job. You do not need all of them. The install name is the value you pass to --skill.
For when the model ships half-finished work: full output, no placeholder comments.
Image generation skills (3)
These produce design images only, no code. Use them with ChatGPT Images, Codex image mode or any agent that generates images, then hand the frames to your coding agent.
Both have three 1 to 10 dials at the top of the file: DESIGN_VARIANCE (centred and clean to asymmetric), MOTION_INTENSITY (hover only to scroll-driven) and VISUAL_DENSITY (spacious to dense). v2 has no dials. Its direction comes from the draw and its limits from the inspector.
Common Questions
What does the default skill need? Node 22 or newer and Chrome, Edge or Chromium. It installs nothing and uses no API keys. It needs the internet only to find pictures.
Where do the pictures come from, and can I use them? From open collections such as Wikimedia Commons, Openverse and the Met. Only public domain, CC0, CC BY and CC BY-SA files are offered, and the skill puts the credits on the page. Replace them with your own when you have them.
Does it work with React, Vue, Svelte, plain HTML? Yes. The rules are about design, not a framework.
Is it for dashboards and app screens? No. The default skill is for public-facing pages. Use minimalist-skill or soft-skill for product UI.
Taste Skill 是一套可移植的 AI 编程助手技能：它能提升 AI 生成界面的水平，让布局、排版、动效和间距更讲究，而不是那种一眼就是模板拼出来的 UI。