Intelligent Web Scraper Agent

A self-aware, self-learning web scraping agent that accumulates experience over time.

Prerequisites

IMPORTANT: Before using this skill, ensure these dependencies are installed.

One-Click Setup (Recommended)

# Navigate to skill directory and run setup
cd ~/.claude/skills/intelligent-web-scraper
chmod +x setup.sh && ./setup.sh

# Or for VPS/headless environment:
./setup.sh --headless

The setup script will:

• Detect your environment (macOS/Linux, GUI/headless)
• Install all Python dependencies
• Install Playwright browsers
• Set up Crawl4AI
• Create necessary directories

Check Dependencies

# Check if everything is installed correctly
python3 ~/.claude/skills/intelligent-web-scraper/scripts/check_deps.py

# Auto-install missing dependencies
python3 ~/.claude/skills/intelligent-web-scraper/scripts/check_deps.py --install

# Output as JSON (for programmatic use)
python3 ~/.claude/skills/intelligent-web-scraper/scripts/check_deps.py --json

Manual Setup (Alternative)

If you prefer manual installation:

# 1. Install Python dependencies
pip install -r ~/.claude/skills/intelligent-web-scraper/requirements.txt

# 2. Install Playwright browser
python -m playwright install chromium

# 3. Set up Crawl4AI
crawl4ai-setup

# 4. For VPS/headless Linux, also run:
python -m playwright install-deps chromium

Required Tools

Tool	Purpose	Installation
Playwright MCP	Browser automation	Must be configured in Claude Code MCP settings
Python 3.9+	Script execution	brew install python or system package manager
Crawl4AI	Advanced crawling	Included in setup.sh

Verify Installation

# Quick verification
python -c "from crawl4ai import AsyncWebCrawler; print('Crawl4AI OK')"
python -c "from playwright.sync_api import sync_playwright; print('Playwright OK')"

# Full check
python3 ~/.claude/skills/intelligent-web-scraper/scripts/check_deps.py

VPS/Headless Deployment

For servers without GUI (VPS, Docker, CI):

# 1. Run setup in headless mode
./setup.sh --headless

# 2. The scraper will automatically use headless browser
# No GUI required - all scraping works via headless Chromium

# 3. Check configuration
python3 ~/.claude/skills/intelligent-web-scraper/scripts/config.py

Note: local_browser_scraper.py (CDP mode) requires a GUI and is not available in headless environments. Use crawl4ai_wrapper.py instead.

---

Local Browser Scraping (CDP)

Use this when the user wants to scrape using their local browser (Comet, Chrome, etc.)

This allows scraping while preserving user's login sessions and cookies!

Quick Start

# 1. Install websockets (one time)
pip install websockets

# 2. Launch Comet/Chrome with debugging (or use the script)
python ~/.claude/skills/intelligent-web-scraper/scripts/local_browser_scraper.py \
    --launch comet \
    --url "https://example.com"

# 3. Scrape data from current page
python ~/.claude/skills/intelligent-web-scraper/scripts/local_browser_scraper.py \
    --extract articles \
    --output data.json

Using local_browser_scraper.py

Launch browser with debugging:

python scripts/local_browser_scraper.py --launch comet --url "https://douban.com"
python scripts/local_browser_scraper.py --launch chrome --url "https://example.com"

Scrape from current tab:

# Get page text
python scripts/local_browser_scraper.py --extract text

# Get all links
python scripts/local_browser_scraper.py --extract links

# Get articles
python scripts/local_browser_scraper.py --extract articles

# Custom JavaScript
python scripts/local_browser_scraper.py --extract "document.querySelectorAll('h1').length"

Scrape specific tab (by URL pattern):

python scripts/local_browser_scraper.py --url-pattern "douban.com" --extract articles

Built-in extractors: text, html, title, links, images, tables, articles, metadata

Manual CDP (Alternative)

CRITICAL: Always use user's EXISTING profile to preserve login sessions!

# Comet - use existing profile
/Applications/Comet.app/Contents/MacOS/Comet \
    --remote-debugging-port=9222 \
    --user-data-dir="$HOME/Library/Application Support/Comet" \
    "https://example.com" &

# Chrome - use existing profile
/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome \
    --remote-debugging-port=9222 \
    --user-data-dir="$HOME/Library/Application Support/Google/Chrome" \
    "https://example.com" &

Browser Profile Locations (macOS)

Browser	User Data Directory
Comet	~/Library/Application Support/Comet
Chrome	~/Library/Application Support/Google/Chrome
Edge	~/Library/Application Support/Microsoft Edge
Brave	~/Library/Application Support/BraveSoftware/Brave-Browser

NEVER use temp directory like /tmp/browser-debug - this creates empty profile without logins!

Why use existing profile:

• Preserves all login sessions (no re-authentication needed)
• Browser extensions work normally
• User's bookmarks, history available

When to Use Local Browser vs Playwright

Use Local Browser	Use Playwright
User requests specific browser	Automated bulk scraping
Need user's login session	Don't need authentication
User wants to see the page	Headless scraping OK
Site blocks headless browsers	Standard sites

---

Core Capabilities

1. Intelligent Page Analysis

• Auto-detect page type (product list, article, series, etc.)
• Identify data containers and selectors
• Recognize pagination mechanisms

2. Smart Pagination & Scroll Loading

• Page numbers, Next button, Infinite scroll, Load more, API pagination
• Auto-detect and handle appropriately

CRITICAL: Scroll Loading Requirements

• MUST automatically scroll down the page until all content is loaded
• Many modern websites use lazy loading/infinite scroll, initially showing only partial content
• Scroll strategy:
  1. Record current page height
  2. Scroll to bottom
  3. Wait 1-2 seconds for new content to load
  4. Check if page height increased
  5. Repeat until height stops changing (3 consecutive times)
• Use browser_evaluate to execute scrolling:

  window.scrollTo(0, document.body.scrollHeight)
  

3. Detail Link Following

CRITICAL: Detail Page Scraping Requirements

• List pages may only show title/summary, full content is on detail pages
• MUST detect and follow detail links to get complete data
• Patterns to identify detail links:
  • "View more", "Read more", "Details", "Full article"
  • Title itself is a link
  • Arrow links like "→"
• Scraping workflow:
  1. First scrape all entries from list page
  2. Identify detail link for each entry
  3. Visit detail pages one by one to extract full content
  4. Merge data and save
• Note: Detail page scraping needs appropriate delays (2-5s) to avoid blocking

4. Anti-Blocking

• Adaptive delays (2-60s based on signals)
• User-Agent rotation
• Block detection and recovery

5. Series Discovery

• Find prev/next links
• Detect table of contents
• URL pattern analysis
• Build complete series from single article

6. Self-Learning (NEW)

• Records successful patterns for each domain
• Accumulates lessons learned
• Reuses knowledge on similar sites

---

Self-Learning System

How It Works

┌─────────────────────────────────────────────────────────────┐
│                    SCRAPE REQUEST                            │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│ Step 1: Check Experience Database                            │
│ - Read experiences/site_patterns.json                        │
│ - Look for matching domain/URL pattern                       │
│ - If found: Use learned selectors and strategies             │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│ Step 2: Execute Scraping                                     │
│ - Use learned patterns OR discover new ones                  │
│ - Apply anti-blocking strategies                             │
│ - Extract data                                               │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│ Step 3: Learn & Record                                       │
│ - If successful: Save patterns to site_patterns.json         │
│ - If failed: Record lesson in lessons_learned.md             │
│ - Update success rate and timestamps                         │
└─────────────────────────────────────────────────────────────┘

Experience Database Structure

Location: ~/.claude/skills/intelligent-web-scraper/experiences/

experiences/
├── site_patterns.json    # Learned patterns for each domain
├── lessons_learned.md    # Accumulated wisdom and gotchas
└── README.md             # How the learning system works

Pattern Storage Format

{
  "domain.com": {
    "url_patterns": {
      "/movie/*/comments": {
        "selectors": {
          "container": ".comment-list",
          "item": ".comment-item",
          "fields": {
            "author": ".comment-info a",
            "content": ".comment-content span",
            "rating": ".rating"
          }
        },
        "pagination": {
          "type": "page_number",
          "selector": ".paginator a"
        },
        "scroll_loading": {
          "required": true,
          "max_scrolls": 20,
          "scroll_delay_ms": 1500,
          "notes": "Page uses infinite scroll loading"
        },
        "detail_links": {
          "required": true,
          "selector": "a.read-more, a:has-text('→')",
          "fields_in_detail": ["fullContent", "images", "comments"],
          "notes": "List page only has titles, need to scrape detail pages for full content"
        },
        "anti_block": {
          "min_delay": 3,
          "max_delay": 8,
          "needs_login": false
        },
        "success_count": 15,
        "last_success": "2024-01-15",
        "notes": "Rate limit after 50 requests"
      }
    }
  }
}

---

Execution Guide

When You Receive a Scraping Request:

Phase 1: Check Existing Knowledge

# ALWAYS do this first
experiences_path = "~/.claude/skills/intelligent-web-scraper/experiences/site_patterns.json"

1. Read site_patterns.json
2. Extract domain from target URL
3. Check if matching pattern exists
4. If yes: Report to user "I have experience with this site, using learned patterns"

Phase 2: Analyze Page

# 1. Navigate to URL
browser_navigate(url)

# 2. Wait for load
browser_wait_for(time=3)

# 3. [IMPORTANT] Scroll to load all content
browser_evaluate('''async () => {
  let prevHeight = 0;
  let currentHeight = document.body.scrollHeight;
  let noChangeCount = 0;

  while (noChangeCount < 3) {
    prevHeight = currentHeight;
    window.scrollTo(0, document.body.scrollHeight);
    await new Promise(r => setTimeout(r, 1500));
    currentHeight = document.body.scrollHeight;

    if (currentHeight === prevHeight) {
      noChangeCount++;
    } else {
      noChangeCount = 0;
    }
  }
  return document.querySelectorAll('article, [class*="item"]').length;
}''')

# 4. Get snapshot (DOM structure)
browser_snapshot()

# 5. Take screenshot for visual analysis
browser_take_screenshot()

Analyze and identify:

• Page Type: product_list | article_list | article_series | single_page
• Data Container: Main wrapper selector
• Data Items: Individual item selector
• Fields: Specific data field selectors
• Pagination: Type and selectors
• Detail Links: Whether detail links need to be followed for more data

Phase 3: Confirm with User

## Page Analysis Report

**URL**: [target]
**Domain Experience**: [Yes, used N times / No, first time]

### Structure Detected
- **Page Type**: [type]
- **Items Found**: ~[N] items on current page
- **Pagination**: [type] ([N] pages estimated)

### Extraction Plan
| Field | Selector | Sample |
|-------|----------|--------|
| title | .item-title | "Example Title" |
| price | .price | "$99.00" |

Continue scraping? [Y/n]

Phase 4: Execute Scraping

Apply appropriate strategy based on pagination type:

• Page Numbers: Iterate through URLs
• Next Button: Click and wait loop
• Infinite Scroll: Scroll and detect new content (see scroll code above)
• Load More: Click button until exhausted

Phase 4.5: Detail Page Scraping

When list page content is incomplete, this step is required:

// 1. Extract all entries and their detail links
const entries = await browser_evaluate(() => {
  const items = document.querySelectorAll('article, .item');
  return Array.from(items).map(item => ({
    title: item.querySelector('h2, h3, .title')?.textContent,
    summary: item.querySelector('p, .desc')?.textContent,
    detailUrl: item.querySelector('a[href*="detail"], a[href*="view"], a:has-text("→")')?.href
  }));
});

// 2. Visit detail pages one by one
for (const entry of entries) {
  if (entry.detailUrl) {
    // Open new tab
    browser_tabs({ action: 'new' });
    browser_navigate(entry.detailUrl);
    browser_wait_for({ time: 2 });

    // Extract detail content
    const detail = await browser_evaluate(() => ({
      fullContent: document.querySelector('article, .content, main')?.innerText,
      images: Array.from(document.querySelectorAll('img')).map(i => i.src),
      metadata: { /* ... */ }
    }));

    entry.detail = detail;

    // Close tab, return to list
    browser_tabs({ action: 'close' });
    await sleep(2000); // Polite delay
  }
}

When detail page scraping is needed:

• List only shows title/date, content is on detail page
• Has "View more", "Read full article" links
• Data fields are obviously incomplete

Phase 5: Learn & Save

CRITICAL: After EVERY successful scrape, update the experience database.

# Update site_patterns.json with:
{
  "selectors": discovered_selectors,
  "pagination": detected_pagination,
  "anti_block": {
    "min_delay": actual_delay_used,
    "max_delay": max_delay_used,
    "blocked_at": request_count_when_blocked or null
  },
  "success_count": previous_count + 1,
  "last_success": today,
  "notes": any_special_observations
}

If scraping fails, add entry to lessons_learned.md:

## [Date] - domain.com - [URL Pattern]
**Issue**: [What went wrong]
**Cause**: [Why it happened]
**Solution**: [How to avoid next time]

---

Anti-Blocking Strategy

Delay Escalation Levels

Level	Delay	Trigger
0 (Normal)	2-5s	Default
1 (Caution)	5-10s	Single 429/503
2 (Careful)	10-20s	Repeated warnings
3 (Critical)	30-60s	Multiple blocks
4 (Pause)	STOP	Captcha detected

Block Signal Detection

BLOCK_SIGNALS = [
    # HTTP Status
    (429, "rate_limited"),
    (403, "forbidden"),
    (503, "service_unavailable"),

    # Page Content
    ("captcha", "captcha_required"),
    ("verification code", "captcha_required"),
    ("access denied", "blocked"),
    ("please wait", "rate_limited"),
]

Recovery Strategy

1. Increase delay exponentially
2. Rotate User-Agent
3. Clear cookies if needed
4. Notify user if captcha detected
5. Record in lessons_learned.md

---

Series Discovery Algorithm

Finding Related Articles

1. Prev/Next Links

   • Look for: prev, next, previous, following, «, »
   • Follow chain in both directions

2. Table of Contents

   • Look for: sidebar TOC, series index, chapter list
   • Extract all links

3. URL Pattern Analysis

   • Detect: /part-1, /chapter-2, ?id=3
   • Generate adjacent URLs

4. Title Similarity

   • Compare titles for series indicators
   • Look for: Part N, Chapter N, (1/10)

Scoring Links

score = (
    url_similarity * 30 +      # Same path structure
    series_indicator * 25 +    # Contains series keywords
    position_weight * 20 +     # In navigation area
    title_similarity * 15 +    # Similar to current title
    same_domain * 10           # Must be same domain
)

# Include if score >= 50

---

Output Formats

Format	Use Case	Extension
JSON	Structured data (products, lists)	.json
Markdown	Article content	.md
CSV	Tabular data	.csv
Auto	Let AI decide based on content	-

---

Example Scenarios

Scenario 1: First-Time Domain

User: Scrape movie reviews from https://movie.douban.com/subject/1292052/comments

Agent:
1. [Check] No existing experience for douban.com/subject/*/comments
2. [Analyze] Opening page, taking screenshot...
   - Detected: Comment list with pagination
   - Container: #comments
   - Items: .comment-item
   - Pagination: Page numbers (.paginator)
3. [Confirm] Found ~20 comments per page, 500+ pages
   Continue? [Y/n]
4. [Execute] Scraping with 3-5s delays...
5. [Learn] Saved pattern to site_patterns.json
   - Next time: Will use learned selectors automatically

Scenario 2: Returning to Known Domain

User: Scrape reviews for another movie https://movie.douban.com/subject/26363254/comments

Agent:
1. [Check] Found experience for douban.com/subject/*/comments
   - Last used: 2024-01-15 (15 successful runs)
   - Known selectors: .comment-item, .comment-content
   - Known rate limit: 3-8s delay recommended
2. [Execute] Using learned patterns directly...
3. [Update] Incremented success_count to 16

Scenario 3: Article Series

User: This is part 3 of a tutorial, get the whole series
      https://blog.example.com/react-tutorial/part-3

Agent:
1. [Analyze] Detected series indicators:
   - Prev link: /react-tutorial/part-2
   - Next link: /react-tutorial/part-4
   - Sidebar TOC: 8 chapters found
2. [Discover] Complete series (8 articles):
   - Part 1: Introduction
   - Part 2: Components
   - Part 3: State (current)
   - ...
3. [Confirm] Scrape all 8 articles?
4. [Execute] Scraping series in order...
5. [Output] Saved to react-tutorial/ directory

Scenario 4: Infinite Scroll + Detail Links

User: Scrape https://v0.app/changelog using Comet browser

Agent:
1. [Check] Found experience for v0.app/changelog
   - scroll_loading: required (infinite scroll)
   - detail_links: some entries have detail pages
2. [Navigate] Opening page in Comet browser...
3. [Scroll] Scrolling to load all content...
   - Scroll 1: 50 articles found
   - Scroll 2: 50 articles (no change)
   - Scroll 3: 50 articles (confirmed complete)
4. [Extract] Extracting list page data...
   - 50 changelog entries extracted
   - 12 entries have detail links (View announcement →)
5. [Detail] Scraping detail pages...
   - Entry 1: "Build with Glean via MCP" → fetching detail...
   - Entry 2: "Linear MCP Integration" → fetching detail...
   - ... (with 2s delay between each)
6. [Save] Merged data saved to v0_changelog.json
7. [Learn] Updated pattern:
   - scroll_loading.max_scrolls: 3 (sufficient)
   - detail_links: most are Twitter links (skip)

---

Resume Capability

Scraping can be interrupted and resumed from where it left off.

How It Works

~/.claude/skills/intelligent-web-scraper/progress/
├── example_com_abc123.json   # Progress for task
├── another_task.json         # Another task
└── ...

Each scraping task automatically saves:

• URLs completed/failed
• Current page number
• Scraped data (incremental)
• Configuration used

Usage

# Start scraping (progress auto-saved)
python scripts/crawl4ai_wrapper.py https://example.com --output data.json

# If interrupted (Ctrl+C), progress is saved
# Resume by running the same command again

# List all progress files
python scripts/progress_manager.py --list

# Show resumable tasks
python scripts/progress_manager.py --resumable

# Check specific task status
python scripts/progress_manager.py --status example_com_abc123

# Delete progress file
python scripts/progress_manager.py --delete example_com_abc123

# Clean up old completed tasks
python scripts/progress_manager.py --cleanup 7  # Delete >7 days old

Programmatic Usage

from progress_manager import ProgressManager

# Start or resume a task
progress = ProgressManager("my_task", start_url="https://example.com")

# Check if resuming
if progress.has_progress():
    print(f"Resuming from {progress.state.completed_urls} URLs")

# During scraping
for url in urls:
    if progress.is_completed(url):
        continue  # Skip already done

    data = scrape(url)
    progress.mark_completed(url, data)

# Finish
progress.finish()

---

Concurrent Scraping

Scrape multiple URLs in parallel with intelligent rate limiting.

Features

• Configurable concurrency: Max parallel requests (global and per-domain)
• Rate limiting: Token bucket algorithm, per-domain limits
• Auto retry: Exponential backoff with jitter
• Progress integration: Automatic resume support
• Graceful shutdown: Ctrl+C saves progress

Usage

# Scrape URLs from file
python scripts/concurrent_scraper.py urls.txt --output results.json

# High concurrency (10 parallel, 3 per domain)
python scripts/concurrent_scraper.py urls.txt -c 10 -d 3

# With rate limiting
python scripts/concurrent_scraper.py urls.txt --rate-limit 5 --domain-rate-limit 1

# Custom delays
python scripts/concurrent_scraper.py urls.txt --min-delay 2 --max-delay 5

Configuration Options

Option	Default	Description
--concurrency, -c	5	Max parallel requests total
--per-domain, -d	2	Max parallel per domain
--rate-limit, -r	10	Global rate limit (req/s)
--domain-rate-limit	2	Per-domain rate limit (req/s)
--min-delay	1.0	Min delay between requests
--max-delay	3.0	Max delay between requests
--retries	3	Max retries per URL
--no-progress	-	Disable resume capability
--task-id	-	Custom task ID for progress

Programmatic Usage

from concurrent_scraper import ConcurrentScraper, ConcurrencyConfig

config = ConcurrencyConfig(
    max_concurrent=5,
    max_per_domain=2,
    global_rate_limit=10.0,
    per_domain_rate_limit=2.0,
)

scraper = ConcurrentScraper(config)
results = await scraper.scrape_urls(url_list)

# Print summary
scraper.print_summary()

Rate Limiting Strategy

┌─────────────────────────────────────────────────────────────┐
│                    Request Flow                              │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
                 ┌────────────────────────┐
                 │   Global Rate Limiter   │  (10 req/s max)
                 │   (Token Bucket)        │
                 └────────────────────────┘
                              │
                              ▼
              ┌───────────────┼───────────────┐
              │               │               │
              ▼               ▼               ▼
       ┌──────────┐    ┌──────────┐    ┌──────────┐
       │ Domain A │    │ Domain B │    │ Domain C │
       │ 2 req/s  │    │ 2 req/s  │    │ 2 req/s  │
       └──────────┘    └──────────┘    └──────────┘
              │               │               │
              ▼               ▼               ▼
       ┌──────────┐    ┌──────────┐    ┌──────────┐
       │ Semaphore│    │ Semaphore│    │ Semaphore│
       │ (max 2)  │    │ (max 2)  │    │ (max 2)  │
       └──────────┘    └──────────┘    └──────────┘

---

Scripts & Archiving

Built-in Scripts

Location: ~/.claude/skills/intelligent-web-scraper/scripts/

Script	Purpose	Usage
crawl4ai_wrapper.py	Advanced Crawl4AI wrapper	<url> --output file.json
concurrent_scraper.py	Concurrent multi-URL scraping	urls.txt -c 5 -o results.json
local_browser_scraper.py	Local browser (Comet/Chrome) CDP scraping	--launch comet --url URL
progress_manager.py	Progress management (resume support)	--list / --resumable
check_deps.py	Dependency checking and installation	--check / --install
config.py	Environment configuration	(displays current config)

Script Archiving Rules

CRITICAL: After each scraping task is completed, the scraping script must be saved in the data output directory!

[data_output_directory]/
├── scrape_[site_name].py    # Scraping script for this site
├── [site_name]_data.json    # Scraped results
└── README.md                # Optional: data description

Reasons:

1. Scripts and data together for easy reuse and traceability
2. Each site has different structure, scripts should be customized
3. Can directly run the script next time for the same site

Script naming convention: scrape_[domain_or_short_name].py

• Example: scrape_v0_changelog.py
• Example: scrape_douban_comments.py

Scripts should include:

1. Target URL
2. Scroll loading logic (if needed)
3. Data extraction logic
4. Output file path

---

Reference Documents

Detailed technical references:

• Crawl4AI API Reference
• Pagination Patterns Guide
• Anti-Blocking Strategies
• Link Discovery Guide

---

Self-Learning Checklist

ALWAYS follow this after every scraping task:

• Read existing site_patterns.json before starting
• Use learned patterns if domain/URL pattern matches
• Scroll check: Does page need scroll loading? Record scroll_loading config
• Detail check: Is list data complete? Need to scrape detail pages? Record detail_links config
• After success: Update site_patterns.json with new/refined patterns
• After failure: Add entry to lessons_learned.md
• Note any special behaviors (rate limits, required delays, gotchas)
• Increment success_count and update last_success date

Scroll loading fields:

• scroll_loading.required: Whether scrolling is needed
• scroll_loading.max_scrolls: How many scrolls to reach bottom
• scroll_loading.scroll_delay_ms: Wait time after each scroll

Detail link fields:

• detail_links.required: Whether detail scraping is needed
• detail_links.selector: Detail link selector
• detail_links.fields_in_detail: Fields only available on detail page

---

Important Notes

1. Respect websites: Follow robots.txt, don't overload servers
2. Legal use only: Only scrape public information
3. Privacy: Don't scrape personal data
4. Self-improve: Always record learnings
5. Share knowledge: Patterns help future scraping tasks