Taste Skill is a set of agent skills for design work: a bundle of art-direction instructions, reference images and a browser verification step that a coding agent loads before it builds a web page. It is aimed at developers who already run Claude Code, Codex or Cursor and are tired of the interchangeable gradient-and-card layouts those agents produce by default. The repository is MIT-licensed, written mostly in JavaScript, and carries 106,172 stars and 7,183 forks on GitHub's own counters. It was created in February 2026 and was last pushed on 30 October 2026, so it is being worked on actively.
What it does
The repository packages taste as instructions an agent can read. Instead of leaving a model to guess what a good page looks like, it hands the agent art direction, reference imagery to work against, and a final pass in a real browser to confirm the result matches.
The most recent release, as the maintainers describe it, is mostly repair work in three areas:
- Markup injection. A chart label built from your own data could inject markup into the page. Labels and headings are now escaped before they reach the screen.
- Accessibility. Every control in the generated output now works from the keyboard and carries an accessible name.
- Typography. The Boska, Erode and Sentient typefaces now load their italic weight correctly.
The maintainers say this whole release came from bugs reported by the community, not from an internal audit.
How it works
The skills sit between you and the agent. You keep using the coding agent you already use, and the skill files supply the design direction the model otherwise invents: what the page should look like, which references to follow, and how the type should be set. The browser check is the part that distinguishes this from a prompt file. The agent opens what it built in an actual browser and verifies it rather than declaring success from the source code alone, which is why the project needs a browser present and not just Node.
Escaping is applied at the point where text is rendered, so data-derived strings such as chart labels and section headings are treated as text instead of markup. The accessibility work is of the same kind: it is applied to what the skill generates, so pages come out keyboard-operable by default rather than being retrofitted later.
Getting started
Two things are required and the project is blunt about them: Node 22, and a real browser. The check does not run without one. No operating system, chip, memory or GPU requirements are listed in the material, so treat it as a normal Node project rather than assuming anything about your machine.
The cost question has a clear answer. The repository is free under MIT, but it is only the instructions: the intelligence that follows them is your coding agent, so you pay whatever Claude Code, Codex or Cursor costs you, by subscription or by API key. Nothing in the material describes a local-model path, so do not plan on one.
The repository is at github.com/Leonxlnx/taste-skill and the project's site is tasteskill.dev. The saved README does not print the install commands, so follow the install section in the repository itself rather than a command quoted second-hand here. The maintainers also publish TasteCode at tastecode.dev, described as Taste Skill packaged as a free and open-source desktop app with a design agent built in.
When to use it / when not
The main catch is the runtime requirement. Node 22 plus a working browser is a real dependency for anyone trying to run this inside a slim CI container or on a locked-down machine, and that requirement did not go away in this release.
Use it instead of just asking Claude for a landing page when you keep rejecting what comes back for looking generic, and you want the same art direction applied every time rather than re-typing it into each session.
Do not use it if you have a house design system and a component library already: the skills exist to supply taste the agent lacks, and if your rules are already codified, you are adding a second opinion you will spend time overriding.
Alternatives
The honest comparison is the one every reader already knows: prompting the agent directly and iterating until the page looks acceptable. That is free and needs no setup, but it is unrepeatable and the browser check is missing. The other option from the same authors is TasteCode, the desktop application, which suits anyone who would rather have an app than skill files inside their agent's configuration.
Developers shipping front ends with a coding agent should take this one seriously, and so should anyone who has already generated pages with an agent and never checked them: a release that escapes data-derived labels and adds keyboard operability is fixing problems that were quietly shipping in agent-built pages. The project is clear about what it needs and clear about where the fixes came from. Just budget for the agent itself, because the repository supplies the taste, not the model.