Web extract

One page in, clean Markdown out — the version an agent can actually reason over.

Method

1. Fetch the URL. Honor redirects; set a real User-Agent. If the response is a PDF, hand off to a document skill instead of guessing.
2. Detect main content by text-density: pick the DOM subtree with the highest ratio of link-free text to markup. Drop <nav>, <aside>, <footer>, cookie and consent banners, share widgets, and comment sections.
3. Convert the surviving tree to Markdown — headings, lists, code blocks, and tables preserved; inline styling flattened; images kept as ![alt](src).
4. Pull metadata from <title>, Open Graph, JSON-LD, and <meta>: title, author, publishedAt, canonical URL, site name.
5. Summarize facts: 3–6 standalone, verifiable statements an agent could cite.

Edge cases

• JS-rendered (empty body, framework shell): report rendered: false so the caller can retry with a headless fetch — don't return an empty article as success.
• Paywalled / truncated: return what's visible and set truncated: true.
• Listing pages (no single article): return top links instead of forcing prose.

Output

{ url, title, byline, publishedAt, siteName, markdown, facts[], wordCount, rendered, truncated }

Never invent metadata: a missing field is null, not a guess.