# Extract DESIGN.md

Turns a domain into a self-contained DESIGN.md — gathering the styleguide, screenshot, and homepage Markdown from context.dev, then composing the document in a single Claude call.

> For the complete documentation index, see [llms.txt](/llms.txt). Markdown variants are available by appending `.md` to any URL or sending an `Accept: text/markdown` header. An agent skill is available at [/.well-known/agent-skills/site-skill.md](/.well-known/agent-skills/site-skill.md).



<DocsBaseSwitcher base="mastra" agent="extract-design-md" />

<AgentPreview
  agent="extract-design-md"
  framework="mastra"
  inputFields="[
  {
    name: &#x22;domain&#x22;,
    label: &#x22;Domain&#x22;,
    placeholder: &#x22;stripe.com&#x22;,
    type: &#x22;text&#x22;,
  },
]"
/>

## Summary [#summary]

The **Extract DESIGN.md Agent** turns any public website into design artifacts.
The `compose_design_md` tool gathers the styleguide, brand, a homepage
screenshot, and clean page Markdown through context.dev, then returns **three
outputs**, which the agent relays verbatim:

* **DESIGN.md** — composed in **one Claude call** (screenshot passed as
  a vision image): token frontmatter (`version: alpha`, colors, typography,
  rounded, spacing, components) plus the canonical sections (Overview, Colors,
  Typography, Layout, Elevation, Shapes, Components, Do's and Don'ts).
* **Tailwind v4 `@theme` block** — derived **deterministically** from the
  styleguide + brand in `lib/derive-tokens.ts`.
* **CSS `:root` tokens** — the same tokens as vanilla CSS variables.

It follows the [designmd.supply](https://designmd.supply) pipeline — the same
prompt spec and deterministic token derivers — running on Claude.

## Installation [#installation]

<CodeTabs>
  <TabsList>
    <TabsTrigger value="cli">
      Command
    </TabsTrigger>

    <TabsTrigger value="manual">
      Manual
    </TabsTrigger>
  </TabsList>

  <TabsContent value="cli">
    ```bash
    npx shadcn@latest add @agentcn/mastra/extract-design-md
    ```
  </TabsContent>

  <TabsContent value="manual">
    <Steps>
      <Step>
        Install the following dependencies:
      </Step>

      ```bash
      npm install @mastra/core context.dev ai @ai-sdk/anthropic
      ```

      <Step>
        Copy and paste the following code into your project.
      </Step>

      <ComponentSource src="registry/mastra/extract-design-md/config.ts" title="config.ts" />

      <ComponentSource src="registry/mastra/extract-design-md/instructions.md" title="instructions.md" />

      <ComponentSource src="registry/mastra/extract-design-md/tools/compose_design_md.ts" title="tools/compose_design_md.ts" />

      <ComponentSource src="registry/mastra/extract-design-md/lib/compose.ts" title="lib/compose.ts" />

      <ComponentSource src="registry/mastra/extract-design-md/lib/design-md.ts" title="lib/design-md.ts" />

      <ComponentSource src="registry/mastra/extract-design-md/lib/derive-tokens.ts" title="lib/derive-tokens.ts" />

      <Step>
        Update the import paths to match your project setup.
      </Step>
    </Steps>
  </TabsContent>
</CodeTabs>

## Composition [#composition]

```text
├── config.ts                  # Agent definition (model + config)
├── instructions.md            # Call compose_design_md, relay all three artifacts
├── tools/
│   └── compose_design_md.ts   # Returns DESIGN.md + Tailwind + CSS
└── lib/
    ├── compose.ts             # Fetches signals; Claude call + token derivers
    ├── design-md.ts           # The verbatim DESIGN.md prompt spec
    └── derive-tokens.ts       # Deterministic Tailwind v4 @theme + CSS :root tokens
```

## Customization [#customization]

* **Data source.** The styleguide, screenshot, and Markdown are pulled through
  [context.dev](https://context.dev). Set `CONTEXT_DEV_API_KEY`, and
  `ANTHROPIC_API_KEY` for the model call.
* **Reshape the document.** Edit `lib/design-md.ts` — the prompt spec is
  the single source of truth for the frontmatter keys, token names, and sections.
* **Swap the model.** Change the model in `lib/compose.ts` — it uses
  `claude-sonnet-4-6` (vision-capable). The reference uses `gpt-5.4-mini`; any
  multimodal model that can read the screenshot works.
* **Temperature.** Adjust `temperature` in `lib/compose.ts` (the reference
  uses `0.2`) to trade determinism for variety.