The Claude Setup That Builds Polished Websites
Set up Claude Code to build polished websites: install the Emil Kowalski, Impeccable, and Taste skills, then connect Figma MCP and Playwright. Full steps.

Claude can now go from a simple idea to a polished, working website without you designing every detail yourself.
The catch is that it does not do this out of the box. Ask any AI model for a landing page with no guidance and you get the same result every time: Inter font, a purple-to-blue gradient, cards inside cards. It works, and it looks like every other AI-built site.
The fix is a five-part setup: three design skills that give Claude better judgment, and two MCP connections that let it design, test, and see its own work.
This Bloh walks through the whole thing with the exact commands.
The stack at a glance
Piece | Type | What it adds |
|---|---|---|
Emil Kowalski’s skills | Skill | Animation and UI polish rules from a design engineer who worked at Vercel and Linear |
Impeccable | Skill | A design vocabulary of 24 commands plus automatic checks for common AI design mistakes |
Taste Skill | Skill | Stronger layout, typography, motion, and spacing in place of boilerplate-looking UI |
Figma MCP | MCP server | A two-way link between Claude and your Figma files |
Playwright MCP | MCP server | A real browser Claude can drive, test in, and screenshot |
Skills change how Claude thinks about design. MCP servers change what Claude can do. You want both.
Before you start
You need:
Claude Code installed and signed in. Follow Anthropic’s official install guide for your operating system.
Node.js, because the skill installers and Playwright MCP run through
npx.A Figma account, for the Figma step.
A project folder. Create one and open your terminal in it, since several installs are scoped to the current project.
mkdir my-website && cd my-websiteStep 1: Install Emil Kowalski’s design skills
Emil Kowalski is a design engineer known for his animation work and the Sonner toast library. His skills list the small mistakes agents make, such as using an ease-in curve on an enter animation or a solid border where a soft shadow belongs, and tell Claude how to avoid them.
Install the whole pack:
npx skills@latest add emilkowalski/skillsOr install only the main skill for Claude Code:
npx skills add emilkowalski/skills --skill emil-design-eng --agent claude-codeThe main skill, emil-design-eng, covers animation decisions and general UI polish. The pack also includes more focused skills, including:
animatefor building an animation from scratch with the right curve and durationreview-animationsfor a strict review of existing motionmobile-nativefor making a web app feel right on a phonepick-ui-libraryso Claude chooses a trusted library over hand-rolling a componentbreak-uifor stress-testing your UI with long names, empty lists, and huge counts
Step 2: Install Impeccable
Impeccable, by Paul Bakaus, builds on Anthropic’s own frontend-design skill. It gives you one skill with 24 commands and a set of deterministic detector rules that catch the usual tells of AI-generated design.
From your project root:
npx impeccable installThe installer detects your tools, asks which to install for, and asks whether to install for this project or globally. Reload Claude Code afterward.
Then, inside Claude Code, run the one-time setup:
/impeccable initThis inspects your project, asks about anything it cannot work out, and writes a PRODUCT.md file describing your audience, purpose, and voice. Later commands read that file, so Claude designs for your product and not for a generic one.
Commands you will use most:
Command | What it does |
|---|---|
| Full plan-then-build flow with visual iteration |
| Plans the UX and UI before any code is written |
| Design review covering hierarchy and clarity |
| Technical checks for accessibility, performance, and responsiveness |
| Fixes fonts, hierarchy, and sizing |
| Fixes layout, spacing, and visual rhythm |
| Adds purposeful motion |
| Final pass before shipping |
To update later, run npx impeccable update.
Step 3: Install Taste Skill
Taste Skill describes itself as an anti-slop frontend framework for AI agents. Its job is to push Claude away from the default AI look and toward layouts, type, and spacing that resemble well-made sites.
Install everything in the repo:
npx skills add <https://github.com/Leonxlnx/taste-skill>Or install only the main skill:
npx skills add <https://github.com/Leonxlnx/taste-skill> --skill "design-taste-frontend"The repo also contains image-generation skills that produce reference boards for web, mobile, and brand kits. You can generate those frames with an image model and hand them to Claude Code as visual direction.
Step 4: Connect Figma MCP
The Figma MCP server lets Claude read your Figma files for design context and work with the canvas. The recommended route in Claude Code is the official Figma plugin, which bundles the MCP connection with Figma’s own skills.
claude plugin install figma@claude-plugins-officialThen:
Restart Claude Code if it was running.
Run
/mcp(or/plugin), select the Figma server, and press Enter to open the authorization page.Sign in to Figma and click Allow access.
If you prefer to add only the MCP server manually:
claude mcp add --transport http figma <https://mcp.figma.com/mcp>Add --scope user to make it available in all your projects. Without that flag it applies only to the current project.
The remote server is link-based: you paste a link to a Figma file, frame, or layer and Claude pulls context from it. Figma also offers a desktop MCP server that runs through the Figma desktop app, but Figma recommends the remote one because it has the broadest feature set.
Step 5: Connect Playwright MCP
Playwright MCP gives Claude a real browser. This is what closes the loop: Claude can open the page it just built, click through it, test it, and take screenshots to check its own work.
claude mcp add playwright npx @playwright/mcp@latestStep 6: Verify everything
Check your MCP servers from the terminal:
claude mcp listYou should see figma and playwright. Inside Claude Code, /mcp shows the same servers and their connection status. Type / to confirm the new skills and the /impeccable command appear.
How to use it: describe the site in plain English
With the setup in place, you no longer need to spell out every section, component, and interaction. You describe the website and let Claude run.
A good first prompt:
Build a landing page for [product], a [one-line description] for [audience]. The tone is [calm / bold / playful]. Sections: hero, three key benefits, social proof, pricing, FAQ, footer. Use the design skills. When you finish each section, open it in the browser with Playwright, take a screenshot at desktop and mobile widths, and fix anything that looks off before moving on.
That last sentence matters most. Without it, Claude writes code and hopes. With it, Claude looks at the result and corrects itself.
A workflow that works well:
Set context once. Run
/impeccable initsoPRODUCT.mdexists.Plan before building. Run
/impeccable shapeon the page you want.Build. Give the plain-English prompt above, or use
/impeccable craft.Let Claude check itself. Ask for Playwright screenshots at several screen widths.
Refine with vocabulary. Use
/impeccable typeset,/impeccable layout, or ask Emil’sreview-animationsskill to review motion.Finish. Run
/impeccable auditand then/impeccable polish.Bring Figma in when needed. Paste a Figma frame link to build from an existing design, or send the built page to Figma for your designer to edit.
Tips for better results
Give references. Name two or three sites you like and say what you like about each. Skills raise the floor, and references set the direction.
Do not load everything on every request. Three design skills overlap. If results feel inconsistent, call one by name for a given task.
Commit often. Use Git so you can roll back when an iteration goes the wrong way.
Be specific about feel. “Fast and quiet, like a developer tool” gives Claude more to work with than “modern and clean.”
Review third-party skills. A skill is a set of instructions your agent follows, and Impeccable also installs project hooks. Skim the
SKILL.mdfiles before you install, as you would with any dependency.
Will this replace front-end developers?
The setup removes a lot of the manual work in going from idea to a good-looking page. That is real, and for a marketing site or an MVP it can take you most of the way.
It does not remove the need for judgment. Emil Kowalski says this himself in his repo: the skills are a side effect of domain expertise, and AI amplifies that expertise instead of replacing it.
Someone still has to decide what the product should say, catch what the screenshots miss, handle accessibility and performance properly, and maintain the code. The developers who learn this stack will simply ship faster than the ones who do not.
Frequently Asked Questions
The commands in this guide are for Claude Code. Skills are portable and can be added to other Claude surfaces, but the claude mcp add and claude plugin install commands are Claude Code features.
Yes, all three repos are public on GitHub and free to install. You still need a Claude plan that includes Claude Code to use them.
No. The design skills and Playwright work without it. Figma MCP is useful if you already have designs or want a designer to edit the result.
No. Start with one, see how the output changes, and then add the next. Impeccable is a good first choice because its setup step captures your product context in a PRODUCT.md file.
Yes. All three skills support several agents, including Cursor and Codex. Both MCP servers work with any MCP-compatible client.
Run npx impeccable update to refresh Impeccable. For the Emil Kowalski and Taste skills, re-run their npx skills add commands.
Playwright MCP gives Claude a real browser it can drive. Claude can open the page it just built, click through it, test it, and take screenshots to check and correct its own work.
No. It removes much of the manual work in going from an idea to a good-looking page, but it does not remove the need for judgment. Someone still has to decide what the product should say, handle accessibility and performance properly, and maintain the code.
Found this useful? You might enjoy this as well

What Is the Modern PM Tech Stack in 2026?
Most PM teams run overlapping AI and prototyping tools. Learn how to build a lean 2026 PM tech stack: fewer tools, one clear job each, and research upstream.

What Is OpenAI Dots? Complete Guide for 2026
Learn what OpenAI dots are, how ChatGPT's always-on AI agents work, who can get one, what they cost, and how to set up your first dot step by step.
