Crawl4ai

Crawl4AI - Website Crawling & Content Extraction

Overview

This skill enables crawling entire websites and extracting clean, structured content with proper metadata separation. It uses an efficient 2-phase approach that separates bulk crawling from content processing.

Two-Phase Architecture

Phase 1: Bulk Crawling (bulk_crawl.py)

• Recursively crawls entire website tree (configurable depth)
• Parallel loading of multiple pages simultaneously
• Excludes nav/header/footer elements during crawl
• Saves raw data: raw.html, raw.md, metadata.json
• Fast and efficient - no per-page processing overhead

Phase 2: Post-Processing (postprocess.py)

• Processes all crawled pages offline
• Cleans markdown content (removes pagination, duplicates, nav remnants)
• Enriches metadata with AI (description, keywords) or heuristic methods
• Saves final: content.md (clean, NO frontmatter!) + enriched metadata.json
• Metadata stays in JSON - markdown stays pure

Key Benefits:

• Separation of Concerns: Crawling ≠ Processing
• Scalability: Crawl entire sites in parallel, process later
• Flexibility: Re-process data without re-crawling
• Clean Output: Markdown without frontmatter, metadata in JSON
• Cost Efficient: Batch AI calls for metadata generation

Workflow

For Single Pages: Use crawl_to_markdown.py for quick single-page extraction with intelligent content filtering.

For Entire Websites (Recommended):

Step 1: Bulk Crawl

python scripts/bulk_crawl.py <url> --output-dir ./site --max-depth 3

• Crawls entire website recursively
• Parallel loading of pages
• Saves raw data for all pages

Step 2: Post-Process

python scripts/postprocess.py ./site

• Cleans all markdown files
• Generates metadata with AI
• Creates final content.md + enriched metadata.json for each page

Result:

site/
  index/
    raw.html          # Original HTML (for reference)
    raw.md            # Original markdown from crawl4ai
    content.md        # ✨ Clean markdown (NO frontmatter!)
    metadata.json     # ✨ All metadata here
  about/
    raw.html
    raw.md
    content.md
    metadata.json
  ...

Crawling Process

Phase 1: Bulk Crawl (bulk_crawl.py)

Crawls an entire website recursively and saves raw data:

python scripts/bulk_crawl.py <url> [options]

Parameters:

• <url>: Start URL to crawl (required)
• --output-dir PATH: Output directory (default: ./crawled_site)
• --max-depth N: Maximum crawl depth (default: 3)
• --wait-time N: JavaScript wait time in seconds (default: 5.0)
• --allow-external: Allow crawling external domains (default: same-domain only)

What it does:

1. Starts from given URL
2. Extracts all internal links
3. Crawls pages in parallel (depth-first)
4. Excludes nav/header/footer during crawl
5. Saves for each page:
   • raw.html - Original HTML
   • raw.md - Raw markdown from crawl4ai
   • metadata.json - Basic metadata (url, title, links, crawled_at)

Examples:

Crawl entire site with max depth 3:

python scripts/bulk_crawl.py https://example.com --max-depth 3

Shallow crawl (only start page + direct links):

python scripts/bulk_crawl.py https://example.com --max-depth 1

Phase 2: Post-Processing (postprocess.py)

Processes bulk-crawled data and generates final output:

python scripts/postprocess.py <crawled_dir> [options]

Parameters:

• <crawled_dir>: Directory with bulk-crawled data (required)
• --no-ai: Disable AI metadata generation (use heuristic methods)

What it does:

1. Finds all pages with raw.md files
2. For each page:
   • Cleans markdown (removes pagination, duplicates, nav remnants)
   • Generates description & keywords (AI or heuristic)
   • Detects language from HTML
   • Calculates content hash and token estimate
   • Saves content.md (clean, NO frontmatter!)
   • Enriches metadata.json with all metadata

Examples:

Post-process with AI (requires ANTHROPIC_API_KEY):

export ANTHROPIC_API_KEY=sk-ant-...
python scripts/postprocess.py ./crawled_site

Post-process without AI (heuristic methods):

python scripts/postprocess.py ./crawled_site --no-ai

Legacy: Single-Page Crawl (crawl_to_markdown.py)

For quick single-page extraction (legacy mode):

python scripts/crawl_to_markdown.py <url> --output-dir ./output

Note: For multi-page sites, use the 2-phase approach (bulk_crawl.py + postprocess.py) instead.

Metadata Structure

Metadata is stored separately in metadata.json (NO frontmatter in markdown!):

{
  "url": "https://example.com/page",
  "crawled_at": "2025-11-04T15:18:23.075744",
  "title": "Page Title",
  "links": ["https://example.com/about", ...],
  "description": "AI-generated or heuristic description (1-2 sentences)",
  "keywords": ["keyword1", "keyword2", ...],
  "language": "en",
  "content_hash": "b2ddd73c87e2af...",
  "estimated_tokens": 1250,
  "processed_at": "2025-11-04T15:18:47.280333"
}

Advantages:

• Clean Separation: Content in .md, metadata in .json
• Easy Processing: Parse JSON without markdown parsing
• Flexible: Add metadata fields without touching content
• Portable: Markdown files work anywhere without frontmatter issues

Content Cleaning

The script automatically removes:

• Navigation menus (main nav, submenu controls)
• Page headers and footers
• Pagination controls (1|2|3, Weiter, Zurück, etc.)
• Slider/carousel controls
• Duplicate sections
• Empty sections and excessive whitespace
• Menu/navigation links

It preserves:

• Main content text and headings
• Content links relevant to the topic
• Images within the main content
• Structured data (lists, tables where present)
• Proper markdown formatting

Output Structure

2-Phase Approach Output

crawled_site/
├── crawl_summary.json    # Summary of crawl (total pages, urls, etc.)
├── index/
│   ├── raw.html          # Original HTML from crawl
│   ├── raw.md            # Original markdown from crawl4ai
│   ├── content.md        # ✨ Clean content (after post-processing)
│   └── metadata.json     # ✨ All metadata (enriched after post-processing)
├── about/
│   ├── raw.html
│   ├── raw.md
│   ├── content.md
│   └── metadata.json
└── contact/
    ├── raw.html
    ├── raw.md
    ├── content.md
    └── metadata.json

File Roles:

• raw.html - Reference/debugging, contains original HTML
• raw.md - Raw output from crawl4ai (before cleaning)
• content.md - Final clean markdown (NO frontmatter!)
• metadata.json - All metadata (title, description, keywords, etc.)

Legacy Single-Page Output

crawled_site/
└── index.md              # Content + frontmatter (legacy mode)

Usage Workflow

Recommended: 2-Phase Workflow

Step 1: Bulk Crawl

# Crawl entire website
python scripts/bulk_crawl.py https://example.com --max-depth 3 --output-dir ./mysite

Step 2: Post-Process

# Process all pages
export ANTHROPIC_API_KEY=sk-ant-...  # Optional, for AI metadata
python scripts/postprocess.py ./mysite

Result: Clean content.md + enriched metadata.json for each page!

Quick Single-Page Extraction (Legacy)

# For single pages only
python scripts/crawl_to_markdown.py https://example.com

Advanced Options

Control crawl depth:

python scripts/bulk_crawl.py https://example.com --max-depth 2  # Shallow crawl

JavaScript-heavy sites:

python scripts/bulk_crawl.py https://example.com --wait-time 10.0

Skip AI metadata generation:

python scripts/postprocess.py ./mysite --no-ai  # Use heuristic methods

KI-gestützte Metadaten (Optional)

Der Skill kann Claude Haiku nutzen, um intelligente Descriptions und Keywords zu generieren.

Mit KI (Empfohlen):

export ANTHROPIC_API_KEY=sk-ant-...

Vorteile:

• Intelligente, kontextbezogene Descriptions
• Semantisch sinnvolle Keywords mit korrekter Großschreibung
• Versteht den Inhalt und fasst ihn zusammen

Kosten: ~0.2 Cent pro Seite (Claude Haiku)

Ohne KI (Fallback):

Wenn kein API-Key gesetzt ist, nutzt der Skill automatisch heuristische Methoden:

• Extrahiert erste substantielle Textzeile als Description
• Ermittelt Keywords nach Häufigkeit
• Funktioniert gut, aber weniger präzise

Der Skill zeigt beim Start an, welcher Modus verwendet wird:

• ✨ KI-Metadaten-Generierung aktiviert (Claude Haiku) - KI wird genutzt
• ℹ️ KI-Metadaten-Generierung nicht verfügbar - Fallback-Methoden werden genutzt

Installation Requirements

The crawl4ai library must be installed in a virtual environment. If the venv doesn't exist yet, create and install:

# Navigate to skill directory
cd /Users/tobiasbrummer/.claude/skills/crawl4ai

# Create virtual environment (if not exists)
python3 -m venv venv

# Activate and install crawl4ai
source venv/bin/activate
pip install crawl4ai

# Install anthropic for KI-based metadata generation (optional but recommended)
pip install anthropic

# Install playwright browsers (one-time setup)
playwright install chromium

Important: All scripts must be run using the venv's Python interpreter:

/Users/tobiasbrummer/.claude/skills/crawl4ai/venv/bin/python3 scripts/crawl_to_markdown.py <url>

See references/crawl4ai_reference.md for detailed API documentation.

Common Issues and Solutions

Issue: Pages taking too long

• Reduce --wait-time if site doesn't need much JavaScript
• Increase --wait-time if content is missing (up to 10s for heavy JS sites)

Issue: Wrong content extracted

• The script uses intelligent heuristics to find main content
• If it picks the wrong element, the site may have unusual structure
• Check the console output to see what selector was used

Issue: Missing content

• Some sites may need longer --wait-time for JavaScript
• Try increasing from 5s to 8-10s for dynamic content
• Check if site requires authentication (not supported)

Issue: Still seeing some navigation

• The 3-stage cleaning is aggressive but may miss some edge cases
• The script will continuously improve pattern matching
• Most navigation is successfully removed (85-95% reduction typical)

Tips for Best Results

1. Default settings work well for most sites
2. Check token estimate in output to verify content size
3. Content hash in frontmatter allows easy change detection for re-crawls
4. Review the extracted content to ensure quality
5. Adjust wait-time for JavaScript-heavy sites
6. Use descriptive output directories when crawling multiple pages