Taste Skill v2: an agent skill that deals a design direction, sources credited photos and fails its own pages in a real browser

One skill with a design and browser-check workflow

The repository contains one MIT-licensed skill, design-taste-frontend, which guides a coding agent through designing and building a site. The current README describes version 2.5: classify the business, plan each screen, deal a palette and layout, build from the supplied system, and check the result in a browser. It supplies assets and scripts for photography, icons, layout, effects and checks.

The finish gate allows up to three rounds of fixes; inspect the resulting page yourself, since automated checks do not guarantee that a design meets the brief.

Three install routes, each with its own update path

The default route uses the npx skills CLI from vercel-labs:

npx skills add https://github.com/Leonxlnx/taste-skill --skill "design-taste-frontend"

Inside Claude Code, the plugin route namespaces the skill under taste-skill:, so it does not collide with other skills you already have installed:

/plugin marketplace add Leonxlnx/taste-skill
/plugin install taste-skill@taste-skill

Upgrades on that route go through /plugin update. For Codex there is a matching plugin route:

codex plugin marketplace add Leonxlnx/taste-skill
codex plugin add taste-skill@taste-skill

Codex only loads the skill in a new session, so an already open session will not see it until you restart. The fourth option is a clone plus ./install.sh, which symlinks the skill into Claude Code, into Codex with --codex, or into both with --all-agents. Because it is a symlink rather than a copy, a git pull updates the installed skill in place. If you already have an older version, running the same install command again is enough, since the install name did not change between versions.

The browser check will not run without Node 22 and a local Chromium

The scripts require Node 22 or newer and an installed Chrome, Edge or Chromium; CHROME_PATH can point to a non-standard installation. The README says the skill installs nothing and needs no API keys. ffmpeg is optional for some image processing.

Bundled Wikimedia Commons photographs work offline. Fetching additional photographs, Google Fonts or brand icons requires network access, so check the assets and font fallbacks in a restricted environment.

Budget depends on the project and model

A full build includes design planning, implementation, browser checks and potentially several rounds of fixes. Time and token usage depend on the model, the brief and the size of the project; the current README does not provide a reliable per-page budget. Measure a representative run before estimating cost.

The workflow is designed for building or redesigning a page, so review whether the full process is proportionate for a small edit.

Photos come from open archives, with credits you have to keep

The current README describes a vetted photo library containing bundled Wikimedia Commons images and Unsplash photographs. The tool provides credits in the page footer and distinguishes draft stand-ins from a client's actual premises, people or products.

Review the source and licence for each chosen asset, including any attribution and reuse conditions, before publication. Replace stand-ins with the client's own photography where appropriate.

The current workflow uses category-specific assets

The current README describes a 2.5 workflow that varies the page by business category and uses supplied UI fragments for software and tech pages. It also describes browser checks for layout, contrast, placeholders and performance, with a finish gate allowing up to three rounds. Its documented viewport lists differ between sections, so consult the scripts for the exact checks in your installed revision.

The supplied starter targets static pages; the README says existing React, Vue, Svelte, Astro and Tailwind projects use their existing stack, while a Next.js brief is built as a static export. Check the project's changelog and pin a revision if repeatable behaviour matters.

Older skill names now point to other repositories

Earlier versions of this repository held many skills. They have been split out, and installing them from this repo no longer works. high-end-visual-design, minimalist-ui, industrial-brutalist-ui, gpt-taste, stitch-design-taste, design-taste-frontend-v1 and design-taste-frontend-v2-experimental now live in Leonxlnx/taste-skill-classics. imagegen-frontend-web, imagegen-frontend-mobile, brandkit and image-to-code moved to Leonxlnx/imagegen-skills. They kept their install names, so only the repository URL changes:

npx skills add https://github.com/Leonxlnx/taste-skill-classics --skill "high-end-visual-design"

Two names do not follow that pattern. redesign-existing-projects was folded into design-taste-frontend. full-output-enforcement was removed outright, and Leonxlnx/unlazy is the suggested replacement for the same problem. If a team script or onboarding doc still installs one of the old names from Leonxlnx/taste-skill, it will not get that skill, so update those references before the next machine is set up.

Editorial conclusion

Use Taste Skill if you run Claude Code, Codex or another agent through the npx skills CLI and want a single designed page or app screen per brief, and you can spend 30 to 45 minutes and a few hundred thousand tokens on each build. Skip it for quick component tweaks or for machines without Node 22+ and Chrome, Edge or Chromium. Before relying on it, confirm that the photo licenses it picks (CC BY, CC BY-SA and others) fit your project, and that you install design-taste-frontend rather than one of the skills that moved to other repos.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes