Teaching Claude Code what good looks like.
Claude Code will write you a settings page, a pricing table and an empty state in one sitting. It will also make all three look like the average of the internet, because a terminal gives it no way to look at anything. Two commands fix that.
The model is not being lazy. It is being blind.
Ask Claude Code to build a checkout and it will produce one — with three fields, a blue button, a reassuring lock icon and a summary card. It does that because that is what most checkouts in its training data look like. The same request from a designer starts with a different question: what do the checkouts that convert actually do, and what do the hostile ones do that I should avoid?
That question needs evidence. A reference library gives it to the agent in the one format it can use mid-task: a tool call that returns named screens, flows, components and licences. See what a UI-reference MCP server is if you want the longer version.
Two commands, in this order.
The key comes from your account page after you sign up. Nothing is installed locally; the server is hosted.
claude mcp add --transport http thinkui \
https://thinkui.xyz/mcp \
--header "Authorization: Bearer your-key"
npx skills add https://thinkui.xyz --skill thinkui -y
{
"mcpServers": {
"thinkui": {
"type": "http",
"url": "https://thinkui.xyz/mcp",
"headers": { "Authorization": "Bearer ${THINKUI_KEY}" }
}
}
}
Verify before you trust it
claude mcp list — the server should appear with a connected status.
/mcp inside a session — prints the servers and the tools each one exposes. Five tools is correct.
A red entry means the URL or command is wrong. A connected entry whose calls fail means the header or the key is wrong, and the error says which.
What to say to it.
The skill makes the tools discoverable; a request that names a decision makes them useful.
Prompts that get evidence
"Before you write the upgrade screen, check how paywalls frame a free trial. Build the calm version."
"Find how good products show an empty search, then use that pattern in the results page."
"What typeface holds up in a dense table? Use one from the library."
What it will not do for you
It will not copy a screenshot into your product. The corpus states each screen's licence, and where it is study-only the result forbids republishing — the agent is told to design from the reasoning, not ship the pixels.
That constraint is deliberate. It is also why results carry a source name: you can go and look at the original.
Questions this page gets.
Why does Claude Code invent interfaces?
It has no visual context of its own. In a terminal there is no design file to open and no browser to compare against, so the model produces the most common pattern in its training data and presents it as a choice.
Does the skill replace the MCP server?
No. The server holds the references; the skill is the habit of asking for them. Without the skill the tools exist but get used rarely; without the server the skill has nothing to call. Install both.
Where does the MCP config live in Claude Code?
Two options. The claude mcp add command writes it for you, or you put the same entry in .mcp.json at the repository root so the whole team inherits it. Project scope needs the file committed; user scope lives in your home directory.
How do I know it is connected?
claude mcp list shows the server with its status. In a session, /mcp prints the connected servers and their tools. If the entry is red the command or URL is wrong; if calls fail after connecting, the Authorization header is missing or the key is wrong.
Does this work in Claude Desktop and claude.ai?
Yes. ThinkUI now supports OAuth 2.1 as well as API keys, so it can be added as a custom connector on claude.ai: paste the URL, sign in, approve access. Claude Desktop and Claude Code take either the OAuth flow or the key, whichever the client offers.
Give it something to look at.
Create an account, run the two commands, and ask for a screen you have opinions about.