Taste Skill (design-taste-frontend) by Leonxlnx for Claude Code
Taste Skill (design-taste-frontend) makes your coding agent design pages with a clear look, real pictures and browser checks, not generic AI layouts.
- Installs
- 716K
- on skills.sh
- Stars
- 107.6K
- on GitHub
- Files
- 38
- besides SKILL.md
- Updated
- Oct 2026
- last repo commit
Taste Skill (install name design-taste-frontend) makes Claude Code, Codex and similar coding agents design web pages and app screens that do not look generic. The agent picks a visual direction, writes the copy first, finds real pictures, then checks the page in a real browser and fixes what looks wrong. Know before you start: it needs Node 22 or newer and Chrome, Edge or Chromium, and a full build takes about 30 to 45 minutes.
Good for
- Developers whose AI-built landing pages all look the same
- Founders shipping a marketing site from a short brief
- Designers who want the agent to follow written house rules
- Teams restyling an existing site without changing the brand
Install design-taste-frontend
Claude Code
- 1
Copy this prompt.
Install the design-taste-frontend skill in Claude Code. Run this command: /plugin marketplace add Leonxlnx/taste-skill /plugin install taste-skill@taste-skill Then check that it loaded and tell me how to use it.
- 2
Paste it into Claude Code. It shows the change and asks before running it.
Requires Node.js 22 or later, Chrome, Edge or Chromium.
Both install the same skill. The browser inspector (inspect.mjs) needs Node 22 or newer and an installed Chrome, Edge or Chromium; draw.mjs, lint.mjs and assets.mjs run on Node 18+. Set CHROME_PATH if your browser is in an unusual place.
From the project README. Source.
What's in the skill
When it loads
tasteskill v2, the anti-slop frontend skill. Art direction for websites, distilled from a study of 247 curated gallery sites (One Page Love, Awwwards Sites of the Day). Deals researched directions (idea, composition, type, palette, photographs, section blueprints) from a library, enforces a strict no-slop house style with a real-browser inspector, and runs a look-and-fix loop on screenshots. Use this whenever you build, design, redesign, restyle, polish or review a landing page, marketing site, portfolio, product page, one-pager, hero or any public-facing web page, in any stack (HTML/CSS, React, Next.js, Vue, Svelte, Astro, Tailwind) — also when the user says a page looks generic, templated, "AI-made", sloppy, or asks to make it look designed, premium, award-level or less like every other site. Also for app screens: dashboards, admin and analytics views, settings, inboxes and other screens people work in, with their own run and rules.
Sections in SKILL.md
- Page or app screen
- The run
- 1. Read the brief and the project
- 2. Write the argument
- 3. Draw the direction
- 4. Find the idea
- 5. Write the words
- 6. Build it, still
- 7. Gate
- 8. Look
- 9. Move
- 10. Report
- House rules
- Laws of the craft
- Special cases
- Files
Key facts
- Author
- Leonxlnx
- License
- MIT
- Installs on skills.sh
- 715.9K (Nov 2, 2026), counts older versions under the same name
- GitHub stars
- 107,778 (Nov 2, 2026)
- Latest version
- 2.1.5 (Oct 25, 2026)
- Last commit on main
- Oct 25, 2026
What it includes
Lists what only this client has
Lists 8 to 12 things only this client has, and what real imagery exists.
Try: “Make a one-page website for a 38-seat Italian restaurant in Lisbon.”
Read ·
Read the briefOne claim and proof per section
Each section gets one claim and its proof. Sections without a claim are deleted.
Write the argumentDeals three of 20 directions to pick from
Deals three researched visual directions. The agent picks one, then gets type, palette, hero, sections, photos and motion.
Runs code ·
draw.mjsWrites all the words before building
Writes the full copy deck first, with short headlines and no invented customers, quotes or stats.
Copy firstFinds credited pictures in open collections
Searches Openverse, Wikimedia Commons, the Art Institute of Chicago, the Met, and video archives, and saves credits.
Files ·
assets.mjsBuilds the page with real pictures
Builds tokens, hero and sections with at least four real pictures, one live interactive element and scroll scenes.
Write ·
BuildScans code for banned patterns and stock phrases
Static scan for banned patterns such as gradients, violet colours and words like seamless.
Runs code ·
lint.mjsChecks the page in a real browser at 3 widths
Renders at 1440, 768 and 390 px, measures the house rules and saves screenshots.
Runs code ·
inspect.mjsPhotographs scroll scenes from 0 to 100%
Captures scroll scenes on desktop, phone and reduced motion.
Runs code ·
scenes.mjsCritiques screenshots and fixes the top three problems
The agent scores screenshots on 14 passes and fixes the top three problems, up to three rounds.
Look and fixAdds required, performance-safe motion
Adds motion using transform and opacity only, with a reduced-motion fallback.
Motion layerGives dashboards their own draw and checks
Dashboards, admin screens and settings use app styles with draw.mjs --app and inspect.mjs --app.
Try: “Build the dashboard for this product with real data and charts.”
App screens routeRestyles a site and keeps your brand
Keeps your stack and brand choices and runs a before and after inspection.
Try: “Redesign this interface while preserving its brand.”
Redesign routeLists direction, scores, picture credits and placeholders
Ends with the direction, the idea sentence, scores, every picture with licence and source, and what to replace.
Report
From the official docs
Permissions
Medium risk
It runs Node scripts and a local browser on your machine and fetches pictures from public collections, and one scanner warns that fetched text can steer the agent. Run it inside a project folder you can recreate and read what it downloads.
What it can do
- Read your data
- Create and change things
- Delete things
- Read-only mode
Watch out for
It starts a local browser and a local web server
The check scripts run Node, launch Chrome, Edge or Chromium and serve your files on 127.0.0.1 for inspection. Review the scripts before the first run, and run the final check on a production build, not a dev server. skills.sh
Web search results go straight to the agent
Snyk flagged that picture searches return third-party text the agent reads, which is a prompt injection risk. Skim what it downloads and keep the credits file it writes. skills.sh
A chart script once trusted unsafe label text
Socket warned that charts.js put untrusted labels into the page. The author fixed this in 2.1.5, so update an older install. github.com
It leaves files outside your project
It keeps a history file in your home folder and temporary browser profiles. Old ad-chrome-* folders from earlier versions are not removed, so delete them by hand. github.com
No keys and no packages to install
It uses no API keys and installs nothing on its own. GitHub lists no security advisories for the repo. github.com
Security advisories
No published security advisories as of Nov 2, 2026.
Technical details
| Turning on read-only | No read-only mode. The lint step only reads files, but the build, picture download and browser check write to the project. |
|---|---|
| Talks to | api.openverse.org, commons.wikimedia.org, api.artic.edu, collectionapi.metmuseum.org, archive.org and images-api.nasa.gov for pictures. fonts.googleapis.com and api.fontshare.com from the pages it builds. cdn.jsdelivr.net (GSAP and Lenis) from pages with scroll scenes |
| Sign-in and keys | No account and no API keys. |
| Data retention | Everything stays on your machine. Screenshots sit in .taste-skill/inspect and are cleared on the next inspector run. |
| Scopes |
|
What people say
People like that it stops generic layouts and dislike how slow and strict it is, and independent proof that it beats no skill is thin.
What people like
- Bans the usual AI tells such as small labels and thin rules; the author counted 37 labels and 249 rules without it and none with it across four briefs
- Lint and a browser check measure the result, which a plain instruction file does not
- Loaded before building, so it prevents generic output instead of flagging it afterwards
Common complaints
- The zero-failure results come from its own inspector, and the author's blind test scored 7.7 with the skill against 7.8 without
- A full build takes 30 to 45 minutes and a few hundred thousand tokens
- Strict house style bans monospace outside code and requires pictures and motion
- One user reported a poor design for a React app built from named third-party components
In their words
“The v2 pages had zero inspector failures, but that inspector is the skill's own rulebook.”DevToolLab TeamHacker NewsAug 2026
“preventive rather than detective. It's a set of SKILL.md files you load into your coding agent (Cursor, Claude Code, etc.) before generation so the slop never gets created in the first place.”Wpnx330GitHub issueOct 2026
“making terrible design...”amazingcreationsltd
Examples
Build a page from a brief with no photos
Make a one-page website for Osteria Nove, a 38-seat Italian restaurant in Alfama, Lisbon. Handmade pasta, a short natural wine list, dinner only, Tuesday to Saturday from 19:00.
The agent lists what only this restaurant has, picks one of three drawn directions, writes the copy, finds credited open-collection pictures and checks the page at three screen widths.
SourceRestyle a site and keep the brand
Redesign this interface while preserving its brand.
The skill takes its redesign route, keeps your stack and brand choices, and compares before and after in the browser check.
SourceDesign a dashboard or app screen
Build the dashboard for this product with real data and charts.
App screens get their own style draw and checks instead of the marketing-page flow.
Source
Questions
How do I install Taste Skill?
Use the install box above. In Claude Code the plugin commands work, and the skills CLI command works in any agent it supports. Make sure Node 22 or newer and Chrome, Edge or Chromium are installed, because the checks need them.
Why is it called design-taste-frontend and not taste-skill?
The repo and folder are taste-skill, but the install name inside the skill is design-taste-frontend. The author keeps both so old links and installs keep working.
Does Taste Skill work in Claude Code, Codex, Cursor and Claude.ai?
It works in Claude Code and Codex, where the README documents it. The author's site also lists Cursor, Gemini CLI and OpenCode. Nobody documents Claude.ai, and the browser checks need a local computer, so do not count on it there.
Is Taste Skill safe?
Mostly. It runs Node scripts and a local browser, and fetches pictures from open collections. On skills.sh, Gen Agent Trust Hub passed it, and Snyk and Socket gave warnings. See the safety section above.
How long does a build take?
About 30 to 45 minutes with the full look-and-fix loop, per the README. Most of that time goes to finding pictures and checking screenshots.
Where did the dials, soft, minimalist and brutalist versions go?
They moved out of this skill. The three dials and the v1 and v2-experimental versions are in Leonxlnx/taste-skill-classics, along with the soft, minimalist, brutalist, GPT and Stitch variants. Image-generation skills moved to Leonxlnx/imagegen-skills. Today's version has no dials and deals a researched direction instead.
Can it build dashboards and app screens?
Yes, since version 2.1. It decides first whether the job is a page or an app screen, and app screens get their own style draw and checks.
Taste Skill vs frontend-design?
frontend-design is Anthropic's single instruction file with nothing that checks the result. Taste Skill adds scripts that pick a direction, find pictures and measure the page, at the cost of time and tokens. Pick frontend-design for quick edits and Taste Skill for a whole page.
Taste Skill vs Impeccable?
Impeccable scans finished HTML, CSS and JSX for design anti-patterns. Taste Skill steers the build before it happens. One Hacker News commenter calls Taste Skill preventive rather than detective, and the two can be used together.
Does it beat using no skill?
The evidence is thin. In the author's own blind test judged by models, the last round scored 7.7 with the skill and 7.8 without, and its builds had no inspector failures against 8 to 13 for the baselines. The inspector is the skill's own rulebook.
Information
updated Nov 2, 2026
- Repository
- leonxlnx/taste-skill
- Folder
skills/taste-skill
Sources
Docs, issue trackers and discussions this page is built from.
Show all 21Hide
- github.com · github.com
- github.com · github.com
- github.com · github.com
- github.com · github.com
- github.com · github.com
- skills.sh · skills.sh
- github.com · github.com
- www.tasteskill.dev · tasteskill.dev
- github.com · github.com
- devtoollab.com · devtoollab.com
- news.ycombinator.com · news.ycombinator.com
- github.com · github.com
- skills.sh · skills.sh
- skills.sh · skills.sh
- skills.sh · skills.sh
- github.com · github.com
- github.com · github.com
- github.com · github.com
- github.com · github.com
- github.com · github.com
- github.com · github.com