Use this when the user wants one URL turned into a practical design system file agents can use immediately.

Default outcome: extract any website's design system in one line and format it as .

The skill should feel like a thin workflow around Firecrawl scrape: gather the page's visible content, structure, metadata, links, and available visual signals, then synthesize those findings into a clean design-system markdown file.

Infer the source URL, target stack, and whether implementation is requested from context. If the user gives a URL and asks for a design system, proceed immediately.

Ask at most 1-3 concise questions only if blocked, such as the website URL, whether to output only or also implement, or a required target stack.

Use the host agent's normal prompt or modal UI. Do not name a harness-specific question function.

Use Firecrawl through the CLI or equivalent tool surface. Always start with two parallel scrapes of the supplied URL:

Example:

firecrawl scrape "https://example.com" --format branding,images -o ".firecrawl/example-branding.json" --pretty &
firecrawl scrape "https://example.com" --full-page-screenshot -o ".firecrawl/example-screenshot.png" &
wait

Combining and in one call still costs a single credit and is required: the block only surfaces curated brand assets (, , , ), so without the agent will miss the page's actual content imagery (heroes, product shots, carousel slides, feature visuals, illustrations, accessory photos, end-of-page artwork, and similar). On a product page like the block has no hero — only returns the main Cybertruck hero (e.g. ) and the rest of the page's photography.

If the screenshot scrape returns a remote image URL (e.g. signed storage link) instead of a local file, download it to the same path so can reference a stable local asset.

Use the structured output as the primary source for colors, typography, components, brand assets (logo, favicon, ogImage), personality, and confidence notes. Use the list as the source of truth for the page's content imagery — hero photography, product shots, carousels, feature visuals, illustrations, and decorative graphics. Use the screenshot as the primary visual reference for layout, hierarchy, and overall feel. Add supplemental formats only when these are insufficient for the final artifact.

Collect:

Do not over-crawl by default. The first version should be useful from a single representative page.

Infer and document the site's design language:

When a value cannot be measured exactly from scrape output, label it as inferred and give a practical approximation.

If appropriate, use sub-agents or equivalent parallel task runners. Natural splits include one page per researcher for multi-page sites, or one reviewer each for colors, typography, spacing, and components.

Each parallel researcher should return source URLs, extracted evidence, inferred design tokens, and confidence notes.

Create or return a with this structure. Embed the full-page screenshot near the top so a coding agent gets visual context alongside the tokens.

undefined

Use this screenshot as the visual source of truth for layout, hierarchy, density, and feel. Tokens below describe the same page in machine-readable form.

[Short description of the visual language and what an agent should recreate]

[Named color roles with hex values when known; mark inferred values clearly]

[Fonts, fallback recommendations, scale, weights, heading/body rules]

[Spacing scale, containers, grids, radius, shadows, borders]

[Buttons, cards, nav, forms, hero, feature sections, pricing, footer, etc.]

[Section order, common layouts, responsive behavior]

[Voice, CTA style, heading patterns, copy density]

[Concrete instructions an AI coding agent can follow to create a new site in this style]

workflow: firecrawl-website-design-clone source_url: [url] target_stack: [stack] output: DESIGN.md

If the user asks to implement, first produce or update `DESIGN.md`, then use it as the source of truth for the build.