Web Scraping

Web scraping methodology

Patterns for reliable, ethical web scraping with fallback strategies and anti-bot handling.

Scraping cascade architecture

Implement multiple extraction strategies with automatic fallback:

from abc import ABC, abstractmethod
from typing import Optional
import requests
from bs4 import BeautifulSoup
import trafilatura

#for .py files
from playwright.sync_api import sync_playwright
from playwright_stealth import stealth_sync

#for .ipynb files
import asyncio
from playwright.async_api import async_playwright

class ScrapingResult:
    def __init__(self, content: str, title: str, method: str):
        self.content = content
        self.title = title
        self.method = method  # Track which method succeeded

class Scraper(ABC):
    @abstractmethod
    def fetch(self, url: str) -> Optional[ScrapingResult]: ...

class TrafilaturaCscraper(Scraper):
    """Fast, lightweight extraction for standard articles."""

    def fetch(self, url: str) -> Optional[ScrapingResult]:
        try:
            downloaded = trafilatura.fetch_url(url)
            if not downloaded:
                return None

            content = trafilatura.extract(
                downloaded,
                include_comments=False,
                include_tables=True,
                favor_recall=True
            )

            if not content or len(content)  Optional[ScrapingResult]:
        import random

        headers = {
            'User-Agent': random.choice(self.USER_AGENTS),
            'Accept': 'text/html,application/xhtml+xml',
            'Accept-Language': 'en-US,en;q=0.9',
        }

        try:
            response = requests.get(url, headers=headers, timeout=30)
            response.raise_for_status()

            soup = BeautifulSoup(response.text, 'html.parser')

            # Remove script/style elements
            for element in soup(['script', 'style', 'nav', 'footer', 'aside']):
                element.decompose()

            # Find main content
            main = soup.find('main') or soup.find('article') or soup.find('body')
            content = main.get_text(separator='\n', strip=True) if main else ''

            title = soup.find('title')
            title_text = title.get_text() if title else ''

            if len(content)  Optional[ScrapingResult]:
        try:
            with sync_playwright() as p:
                browser = p.chromium.launch(headless=True)
                context = browser.new_context(
                    viewport={'width': 1920, 'height': 1080},
                    user_agent='Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36'
                )
                page = context.new_page()

                # Apply stealth to avoid detection
                stealth_sync(page)

                page.goto(url, wait_until='networkidle', timeout=60000)

                # Wait for content to load
                page.wait_for_timeout(2000)

                # Extract content
                content = page.evaluate('''() => {
                    const article = document.querySelector('article, main, .content, #content');
                    return article ? article.innerText : document.body.innerText;
                }''')

                title = page.title()

                browser.close()

                if len(content)  Optional[ScrapingResult]:
        try:
            async with async_playwright() as p:
                browser = await p.chromium.launch(headless=True)
                context = await browser.new_context(
                    viewport={'width': 1920, 'height': 1080},
                    user_agent='Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36'
                )
                page = await context.new_page()

                # Note: playwright-stealth async version
                # from playwright_stealth import stealth_async
                # await stealth_async(page)

                await page.goto(url, wait_until='networkidle', timeout=60000)

                # Wait for content to load
                await page.wait_for_timeout(2000)

                # Extract content
                content = await page.evaluate('''() => {
                    const article = document.querySelector('article, main, .content, #content');
                    return article ? article.innerText : document.body.innerText;
                }''')

                title = await page.title()

                await browser.close()

                if len(content)  Optional[ScrapingResult]:
        for scraper in self.scrapers:
            result = scraper.fetch(url)
            if result:
                return result
        return None

Anti-bot landscape (as of 2026-05)

The cascade above (requests → trafilatura → Playwright + playwright-stealth) handles plain HTML and lightly-protected JS sites. Modern anti-bot stacks (Cloudflare Bot Management / Turnstile, DataDome, Akamai Bot Manager, PerimeterX) layer multiple detection signals: TLS / HTTP-2 fingerprints, browser fingerprints, JS-execution proofs, residential-IP reputation, session behavior. No single tool defeats all of them.

playwright-stealth (2.0+, current) patches obvious detection vectors — navigator.webdriver, chrome.runtime, plugin enumeration, language settings, WebGL fingerprints. Treat it as the floor, not the ceiling. If a target fingerprints TLS or runs Turnstile, stealth alone won't pass.

| Tool | Layer it addresses | Notes | |---|---|---| | curl_cffi | TLS / HTTP-2 fingerprint | Drop-in replacement for requests that mimics Chrome/Safari/Edge JA3+ALPN. Can't run JS — pair with a parsed-HTML extractor when JS isn't required. | | playwright-stealth 2.x | JS-runtime fingerprint | The starting line for Playwright/Chromium. Updates lag the bot stacks; expect to combine with rotation. | | Camoufox | JS + browser fingerprint at C++ level | Firefox-based stealth browser. Spoofs fingerprint values low enough that JS-side checks can't see through them. Use when Chromium-based stealth is detected. | | SeleniumBase UC Mode | Turnstile + browser fingerprint | The closest thing to a one-shot Turnstile solver in 2026, but heavier than playwright-stealth. | | Residential proxy pool | IP reputation | Datacenter IPs (DigitalOcean, AWS) get challenged on first request. Residential pools cost more but bypass the cheapest layer of defense. |

Use the lightest tool that works. Targets without aggressive defense don't need Camoufox or proxy pools — curl_cffi plus a sleep is usually enough. Reserve heavier tools for sites that explicitly serve a Turnstile challenge or DataDome interstitial.

Undocumented APIs

Finding undocumented APIs

Use browser developer tools to discover APIs:

1. Open developer tools (right-click → Inspect, or F12)
2. Go to the Network tab to monitor all requests
3. Filter by Fetch/XHR to show only API calls
4. Trigger the action you want to capture (search, scroll, click)
5. Analyze the response — usually JSON with key-value pairs
6. Copy as cURL (right-click the request)
7. Convert to code using curlconverter.com

Stripping down API requests

When you copy a cURL from dev tools, it includes many parameters. Strip it down by:

1. Remove unnecessary cookies — test without them first
2. Keep authentication tokens if required
3. Identify the input parameters you can modify (like prefix for search terms)
4. Test parameter values — some expire, so periodically verify

Example: Reverse-engineering an autocomplete API

import requests
import time

def search_suggestions(keyword: str) -> dict:
    """
    Get autocompleted search suggestions from an undocumented API.
    Stripped down from browser dev tools capture.
    """
    headers = {
        'User-Agent': 'Mozilla/5.0 (Macintosh; Intel Mac OS X 10.15; rv:100.0) Gecko/20100101 Firefox/100.0',
        'Accept': 'application/json, text/javascript, */*; q=0.01',
        'Accept-Language': 'en-US,en;q=0.5',
    }

    params = {
        'prefix': keyword,
        'suggestion-type': ['WIDGET', 'KEYWORD'],
        'alias': 'aps',
        'plain-mid': '1',
    }

    response = requests.get(
        'https://completion.amazon.com/api/2017/suggestions',
        params=params,
        headers=headers
    )
    return response.json()

# Collect suggestions for multiple keywords
keywords = ['a', 'b', 'cookie', 'sock']
data = []

for keyword in keywords:
    suggestions = search_suggestions(keyword)
    suggestions['search_word'] = keyword  # track seed keyword
    time.sleep(1)  # rate limit yourself
    data.extend(suggestions.get('suggestions', []))

Source: Leon Yin, "Finding Undocumented APIs," Inspect Element, 2023

Poison pill detection

Detect paywalls, anti-bot pages, and other failures:

from dataclasses import dataclass
from enum import Enum
import re

class PoisonPillType(Enum):
    PAYWALL = 'paywall'
    CAPTCHA = 'captcha'
    RATE_LIMIT = 'rate_limit'
    CLOUDFLARE = 'cloudflare'
    LOGIN_REQUIRED = 'login_required'
    NOT_FOUND = 'not_found'
    NONE = 'none'

@dataclass
class PoisonPillResult:
    detected: bool
    type: PoisonPillType
    confidence: float
    details: str

class PoisonPillDetector:
    PATTERNS = {
        PoisonPillType.PAYWALL: [
            r'subscribe to continue',
            r'subscription required',
            r'become a member',
            r'sign up to read',
            r'you\'ve reached your limit',
            r'article limit reached',
        ],
        PoisonPillType.CAPTCHA: [
            r'verify you are human',
            r'captcha',
            r'robot verification',
            r'prove you\'re not a robot',
        ],
        PoisonPillType.RATE_LIMIT: [
            r'too many requests',
            r'rate limit exceeded',
            r'slow down',
            r'429',
        ],
        PoisonPillType.CLOUDFLARE: [
            r'checking your browser',
            r'cloudflare',
            r'ddos protection',
            r'please wait while we verify',
        ],
        PoisonPillType.LOGIN_REQUIRED: [
            r'sign in to continue',
            r'log in required',
            r'create an account',
        ],
    }

    PAYWALL_DOMAINS = {
        'nytimes.com': PoisonPillType.PAYWALL,
        'wsj.com': PoisonPillType.PAYWALL,
        'washingtonpost.com': PoisonPillType.PAYWALL,
        'ft.com': PoisonPillType.PAYWALL,
        'bloomberg.com': PoisonPillType.PAYWALL,
    }

    def detect(self, url: str, content: str, status_code: int = 200) -> PoisonPillResult:
        # Check status code
        if status_code == 429:
            return PoisonPillResult(True, PoisonPillType.RATE_LIMIT, 1.0, 'HTTP 429')
        if status_code == 403:
            return PoisonPillResult(True, PoisonPillType.CLOUDFLARE, 0.8, 'HTTP 403')
        if status_code == 404:
            return PoisonPillResult(True, PoisonPillType.NOT_FOUND, 1.0, 'HTTP 404')

        # Check known paywall domains
        from urllib.parse import urlparse
        domain = urlparse(url).netloc.replace('www.', '')
        for paywall_domain, pill_type in self.PAYWALL_DOMAINS.items():
            if paywall_domain in domain:
                # Check if content is suspiciously short (paywall truncation)
                if len(content)  dict:
    """Extract metadata without downloading video."""
    ydl_opts = {
        'skip_download': True,
        'quiet': True,
        'no_warnings': True,
    }

    with yt_dlp.YoutubeDL(ydl_opts) as ydl:
        info = ydl.extract_info(url, download=False)
        return {
            'title': info.get('title'),
            'description': info.get('description'),
            'duration': info.get('duration'),
            'upload_date': info.get('upload_date'),
            'view_count': info.get('view_count'),
            'channel': info.get('channel'),
            'thumbnail': info.get('thumbnail'),
        }

def download_video(url: str, output_dir: Path, audio_only: bool = False) -> Path:
    """Download video or audio."""
    output_template = str(output_dir / '%(title)s.%(ext)s')

    ydl_opts = {
        'outtmpl': output_template,
        'quiet': True,
    }

    if audio_only:
        ydl_opts['format'] = 'bestaudio/best'
        ydl_opts['postprocessors'] = [{
            'key': 'FFmpegExtractAudio',
            'preferredcodec': 'mp3',
        }]

    with yt_dlp.YoutubeDL(ydl_opts) as ydl:
        info = ydl.extract_info(url, download=True)
        filename = ydl.prepare_filename(info)
        if audio_only:
            filename = filename.rsplit('.', 1)[0] + '.mp3'
        return Path(filename)

def get_transcript(url: str) -> list[dict]:
    """Extract auto-generated or manual subtitles."""
    ydl_opts = {
        'skip_download': True,
        'writesubtitles': True,
        'writeautomaticsub': True,
        'subtitleslangs': ['en'],
        'quiet': True,
    }

    with yt_dlp.YoutubeDL(ydl_opts) as ydl:
        info = ydl.extract_info(url, download=False)

        # Check for subtitles
        subtitles = info.get('subtitles', {})
        auto_captions = info.get('automatic_captions', {})

        # Prefer manual subtitles over auto-generated
        subs = subtitles.get('en') or auto_captions.get('en')
        if not subs:
            return []

        # Get the vtt or json format
        for sub in subs:
            if sub['ext'] in ['vtt', 'json3']:
                # Download and parse subtitle file
                # ... implementation depends on format
                pass

        return []

Instagram with instaloader

import instaloader
from pathlib import Path

class InstagramScraper:
    def __init__(self, username: str = None, session_file: str = None):
        self.loader = instaloader.Instaloader(
            download_videos=True,
            download_video_thumbnails=False,
            download_geotags=False,
            download_comments=False,
            save_metadata=True,
            compress_json=False,
        )

        if session_file and Path(session_file).exists():
            self.loader.load_session_from_file(username, session_file)

    def get_profile_posts(self, username: str, limit: int = 50) -> list[dict]:
        """Get recent posts from a profile."""
        profile = instaloader.Profile.from_username(self.loader.context, username)
        posts = []

        for i, post in enumerate(profile.get_posts()):
            if i >= limit:
                break

            posts.append({
                'shortcode': post.shortcode,
                'url': f'https://instagram.com/p/{post.shortcode}/',
                'caption': post.caption,
                'timestamp': post.date_utc.isoformat(),
                'likes': post.likes,
                'comments': post.comments,
                'is_video': post.is_video,
                'video_url': post.video_url if post.is_video else None,
            })

        return posts

    def download_post(self, shortcode: str, output_dir: Path):
        """Download a single post's media."""
        post = instaloader.Post.from_shortcode(self.loader.context, shortcode)
        self.loader.download_post(post, target=str(output_dir))

TikTok with yt-dlp

def scrape_tiktok_profile(username: str, output_dir: Path, limit: int = 50) -> list[dict]:
    """Scrape TikTok profile videos."""
    profile_url = f'https://tiktok.com/@{username}'

    ydl_opts = {
        'quiet': True,
        'extract_flat': True,  # Don't download, just get info
        'playlistend': limit,
    }

    with yt_dlp.YoutubeDL(ydl_opts) as ydl:
        info = ydl.extract_info(profile_url, download=False)
        videos = []

        for entry in info.get('entries', []):
            videos.append({
                'id': entry.get('id'),
                'title': entry.get('title'),

…

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [jamditis](https://github.com/jamditis)
- **Source:** [jamditis/claude-skills-journalism](https://github.com/jamditis/claude-skills-journalism)
- **License:** MIT
- **Homepage:** https://skills.amditis.tech/

Install and usage instructions live in the source repository linked above.