Cloudflare Markdown Conversion

Use this skill to convert URLs or local files (PDFs, Images, HTML, CSV, Office docs) into clean, structured Markdown for text analysis, RAG, and LLMs.

Features & Supported Formats

• Scraping URLs: Extracts HTML, resolves relative links, handles JSON-LD, extracts title/description.
• Images: Automatically runs object-detection and uses an LLM (gemma-3-12b-it) to generate image descriptions. Converts SVG to raster.
• PDFs: Parses internal StructTree tagging for high-fidelity semantic Markdown extraction.
• Office Docs: Supports .docx, .xlsx, .csv, .ods, .odt, and more.

Usage

Setup & Authentication

This skill requires CLOUDFLARE_ACCOUNT_ID and CLOUDFLARE_API_TOKEN.

Automatic Setup: For convenience, the script automatically looks for a .env file in the current directory or its parents. If you have a .env file in your project root, it will work out of the box.

Manual Setup: Alternatively, you can export them in your shell or pass them as parameters (--account and --token).

Instruction for the Agent: If the skill fails due to missing environment variables, check if a .env file exists in the workspace root.

Scraping a URL

# Basic usage (defaults to 'auto' method, trying AI parsing first, then browser rendering)
node scripts/render.js --url "https://example.com"

Scraping with Options (CSS Selectors, etc.)

Cloudflare allows filtering elements using cssSelector or providing a hostname.

# Only extract the main content container
node scripts/render.js --url "https://developer.cloudflare.com" \
  --options '{"html": {"cssSelector": "main.content"}}'

Converting a Local File (PDFs, Images, Office Docs)

node scripts/render.js --file "report.pdf"

Converting Images with Language Options

Image descriptions are generated via AI. You can specify a desired output language for the description (en, it, de, es, fr, pt).

node scripts/render.js --file "cat.jpeg" \
  --options '{"image": {"descriptionLanguage": "es"}}'

Advanced Options for JS-Heavy Sites

If a site requires complex JavaScript rendering or redirects, use the browser method with specific wait conditions.

# Wait for network to be idle before extracting content
node scripts/render.js --url "https://complex-site.com" --wait "networkidle2"

# Wait for a specific element to appear (e.g. price or main content)
node scripts/render.js --url "https://shop.com/prod" --selector ".product-price"

# Increase timeout for slow pages (in milliseconds)
node scripts/render.js --url "https://slow-site.com" --timeout 60000

Valid --wait options are: load, domcontentloaded (default), networkidle0, and networkidle2.

How It Works Intelligently

The --method auto capability tests two separate rendering paths:

1. Workers AI tomarkdown (Primary): Ideal for documents, standard web pages, extracting JSON-LD structured data, and resolving standard HTML features. Uses multipart form data.
2. Browser Rendering API (Fallback): If the page uses complex JavaScript (e.g. Single Page Apps) and the AI path cannot see the content, the Browser Rendering engine opens a headless real browser for accurate conversion.

Calling the REST API Directly (Advanced)

If you'd prefer not to use scripts/render.js, here is the curl equivalent for a local file using the tomarkdown REST API:

curl https://api.cloudflare.com/client/v4/accounts/${CLOUDFLARE_ACCOUNT_ID}/ai/tomarkdown \
  -X POST \
  -H "Authorization: Bearer ${CLOUDFLARE_API_TOKEN}" \
  -F "files=@document.pdf" \
  -F 'conversionOptions={"pdf":{"metadata":false}}'

Note: For URLs, you should use curl to fetch the source to a local file first before uploading it as files=@<temp.html>. The tomarkdown REST API does not directly ingest a --data url="https...".