| name | deep-read |
| description | Read and extract content from JS-rendered web pages using headless Chrome (Cloudflare Browser Rendering). Unlike /summarize, handles SPAs, paywalled previews, and complex sites. Use when a URL fails with regular fetch, or when user says "read this page", "what does this page say", "get the content from this URL".
|
| allowed-tools | ["mcp__cf-browser__browser_markdown","mcp__cf-browser__browser_screenshot","mcp__cf-browser__browser_json","mcp__cf-browser__browser_links","WebFetch","Read","Write"] |
| model | haiku |
| user-invocable | true |
| when_to_use | When user shares a URL from a JS-heavy site (SPA, dashboard, news behind JS) or when /summarize returns garbled/empty content. Requires CF credentials.
|
| argument-hint | <URL> [extract: summary|data|links|screenshot] |
Deep Read
You are the deep page reading skill for Claude-Agent. Extract content from web pages that regular fetching can't handle.
Input
$ARGUMENTS — a URL and optional extraction mode (default: summary).
Steps
1. Check Availability
If cf-browser MCP is unavailable, fall back to WebFetch with a note:
"Using basic fetch — for JS-heavy pages, set up cf-browser (see README)."
2. Extract Content
Based on extraction mode:
summary (default):
- Call
browser_markdown(url=<URL>) — returns clean markdown of the rendered page
- Summarize: TL;DR (1-2 sentences) + Key Points (3-5 bullets)
data:
- Call
browser_json(url=<URL>, prompt="Extract the main structured data") — returns JSON
- Present as a formatted table or structured output
links:
- Call
browser_links(url=<URL>) — returns all hyperlinks on the page
- Group by type: navigation, content, external
screenshot:
- Call
browser_screenshot(url=<URL>) — returns PNG screenshot
- Save to
workspace/screenshot-[domain].png
3. Save if Valuable
Save extracted content to workspace/deepread-[domain-slug].md.
4. Present
Format the output concisely:
- For summaries: TL;DR + Key Points (under 200 words)
- For data: formatted table
- For links: grouped list
- For screenshots: file path
Rules
- Always try
browser_markdown first — it's the cleanest output
- If the page requires login, note this and suggest the user provide cookies
- Keep summaries under 200 words
- Use the user's language for the summary
