Web Scraping

Kanitsal Cerceve (Evidential Frame Activation)

Kaynak dogrulama modu etkin.

[assert|neutral] Systematic web data extraction workflow with sequential-thinking planning phase [ground:skill-design:2026-01-12] [conf:0.85] [state:provisional]

Overview

Web scraping enables structured data extraction from web pages through the claude-in-chrome MCP server. This skill enforces a PLAN-READ-TRANSFORM pattern focused exclusively on data extraction without page modifications.

Philosophy: Data extraction fails when executed without understanding page structure. By mandating structure analysis before extraction, this skill improves data quality and reduces selector errors.

Methodology: Six-phase execution with emphasis on READ operations:

1. PLAN Phase: Sequential-thinking MCP analyzes extraction requirements
2. NAVIGATE Phase: Navigate to target URL(s)
3. ANALYZE Phase: Understand page structure via read_page
4. EXTRACT Phase: Pull data using get_page_text and javascript_tool
5. TRANSFORM Phase: Convert to structured format (JSON/CSV/Markdown)
6. STORE Phase: Persist to Memory MCP

Key Differentiator from browser-automation:

• READ-ONLY focus (no form submissions, no button clicks for actions)
• Emphasis on data transformation and output formats
• Pagination handling for multi-page datasets
• Rate limiting awareness for bulk extraction
• No account creation, no purchases, no write operations

Value Proposition: Transform unstructured web content into clean, structured datasets suitable for analysis, reporting, or system integration.

When to Use This Skill

Trigger Thresholds:

Primary Use Cases:

• Product catalog extraction (prices, descriptions, images)
• Article content scraping (headlines, body text, metadata)
• Table data extraction (financial data, statistics, listings)
• Directory scraping (contact info, business listings)
• Research data collection (citations, abstracts, datasets)
• Price monitoring (competitive analysis, market research)

Apply When:

• Task requires structured output (JSON, CSV, Markdown table)
• Multiple pages need to be traversed (pagination)
• Data needs normalization or transformation
• Consistent data schema is required
• Historical data collection for trend analysis

When NOT to Use This Skill

• Form submissions or account actions (use browser-automation)
• Interactive workflows requiring clicks (use browser-automation)
• Single-page simple text extraction (use get_page_text directly)
• API-accessible data (use direct API calls instead)
• Real-time streaming data (use WebSocket or API)
• Content behind authentication requiring login (use browser-automation first)
• Sites with explicit anti-scraping measures (respect robots.txt)

Core Principles

Principle 1: Plan Before Extract

Mandate: ALWAYS invoke sequential-thinking MCP before data extraction.

Rationale: Web pages have complex DOM structures. Planning reveals data locations, identifies patterns, and anticipates pagination before extraction begins.

In Practice:

• Analyze page structure via read_page first
• Map data field locations to selectors
• Identify pagination patterns
• Plan output schema before extraction
• Define data validation rules

Principle 2: Read-Only Operations

Mandate: NEVER modify page state during extraction.

Rationale: Web scraping is about data retrieval, not interaction. Write operations (form submissions, clicks that change state) belong to browser-automation.

Allowed Operations:

• navigate (to target URLs)
• read_page (DOM analysis)
• get_page_text (text extraction)
• javascript_tool (read-only DOM queries)
• computer screenshot (visual verification)
• scroll (for lazy-loaded content)

Prohibited Operations:

• form_input (except for search filters if needed)
• computer left_click on submit buttons
• Any action that modifies page state
• Account creation or login actions

Principle 3: Structured Output First

Mandate: Define output schema before extraction begins.

Rationale: Knowing the target format guides extraction strategy and ensures consistent data quality across all pages.

Output Formats:

Principle 4: Pagination Awareness

Mandate: Detect and handle pagination patterns before starting extraction.

Rationale: Most valuable datasets span multiple pages. Failing to handle pagination yields incomplete data.

Pagination Patterns:

Principle 5: Rate Limiting Respect

Mandate: Implement delays between requests to avoid overwhelming servers.

Rationale: Aggressive scraping can trigger blocks, harm server performance, and violate terms of service.

Guidelines:

Implementation:

// Use computer wait action between page loads
await mcp__claude-in-chrome__computer({
  action: "wait",
  duration: 2,  // seconds
  tabId: tabId
});

Production Guardrails

MCP Preflight Check Protocol

Before executing any web scraping workflow, run preflight validation:

Preflight Sequence:

async function preflightCheck() {
  const checks = {
    sequential_thinking: false,
    claude_in_chrome: false,
    memory_mcp: false
  };

  // Check sequential-thinking MCP (required)
  try {
    await mcp__sequential-thinking__sequentialthinking({
      thought: "Preflight check - verifying MCP availability for web scraping",
      thoughtNumber: 1,
      totalThoughts: 1,
      nextThoughtNeeded: false
    });
    checks.sequential_thinking = true;
  } catch (error) {
    console.error("Sequential-thinking MCP unavailable:", error);
    throw new Error("CRITICAL: sequential-thinking MCP required but unavailable");
  }

  // Check claude-in-chrome MCP (required)
  try {
    const context = await mcp__claude-in-chrome__tabs_context_mcp({});
    checks.claude_in_chrome = true;
  } catch (error) {
    console.error("Claude-in-chrome MCP unavailable:", error);
    throw new Error("CRITICAL: claude-in-chrome MCP required but unavailable");
  }

  // Check memory-mcp (optional but recommended)
  try {
    checks.memory_mcp = true;
  } catch (error) {
    console.warn("Memory MCP unavailable - extracted data will not be persisted");
    checks.memory_mcp = false;
  }

  return checks;
}

Error Handling Framework

Error Categories:

Error Recovery Pattern:

async function extractWithRetry(extractor, context, maxRetries = 3) {
  let lastError = null;

  for (let attempt = 1; attempt <= maxRetries; attempt++) {
    try {
      const data = await extractor(context);
      return data;
    } catch (error) {
      lastError = error;
      console.error(`Extraction attempt ${attempt} failed:`, error.message);

      if (isRateLimitError(error)) {
        const delay = Math.pow(2, attempt) * 5; // 10s, 20s, 40s
        await sleep(delay * 1000);
      } else if (!isRecoverableError(error)) {
        break;
      }
    }
  }
  throw lastError;
}

function isRecoverableError(error) {
  const nonRecoverable = [
    "CRITICAL: sequential-thinking MCP required",
    "CRITICAL: claude-in-chrome MCP required",
    "Access denied",
    "CAPTCHA",
    "Blocked"
  ];
  return !nonRecoverable.some(msg => error.message.includes(msg));
}

Data Validation Framework

Purpose: Ensure extracted data meets quality standards before storage.

Validation Rules:

const VALIDATION_RULES = {
  required_fields: ["name", "url"],  // Must be present
  field_types: {
    price: "number",
    name: "string",
    url: "url",
    date: "date"
  },
  constraints: {
    price: { min: 0, max: 1000000 },
    name: { minLength: 1, maxLength: 500 },
    url: { pattern: /^https?:\/\// }
  }
};

function validateRecord(record, rules) {
  const errors = [];

  // Check required fields
  for (const field of rules.required_fields) {
    if (!record[field]) {
      errors.push(`Missing required field: ${field}`);
    }
  }

  // Check field types
  for (const [field, type] of Object.entries(rules.field_types)) {
    if (record[field] && !isValidType(record[field], type)) {
      errors.push(`Invalid type for ${field}: expected ${type}`);
    }
  }

  return { valid: errors.length === 0, errors };
}

Main Workflow

Phase 1: Planning (MANDATORY)

Purpose: Analyze extraction requirements and define data schema.

Process:

1. Invoke sequential-thinking MCP
2. Define target data fields and types
3. Plan selector strategy
4. Identify pagination pattern
5. Define output format
6. Plan rate limiting strategy

Planning Questions:

• What data fields need to be extracted?
• What is the expected output format (JSON/CSV/Markdown)?
• Is pagination involved? What pattern?
• What validation rules apply?
• What rate limiting is appropriate?

Output Contract:

extraction_plan:
  target_url: string
  output_format: "json" | "csv" | "markdown"
  schema:
    fields:
      - name: string
        selector: string
        type: string | number | url | date
        required: boolean
  pagination:
    type: "numbered" | "next_button" | "infinite_scroll" | "load_more" | "none"
    max_pages: number
    delay_seconds: number
  rate_limit:
    delay_between_requests: number

Phase 2: Navigation

Purpose: Navigate to target URL and establish context.

Process:

1. Get tab context (tabs_context_mcp)
2. Create new tab for scraping (tabs_create_mcp)
3. Navigate to starting URL
4. Take initial screenshot for verification
5. Wait for page load completion

Implementation:

// 1. Get existing context
const context = await mcp__claude-in-chrome__tabs_context_mcp({});

// 2. Create dedicated tab
const newTab = await mcp__claude-in-chrome__tabs_create_mcp({});
const tabId = newTab.tabId;

// 3. Navigate to target
await mcp__claude-in-chrome__navigate({
  url: targetUrl,
  tabId: tabId
});

// 4. Wait for page load
await mcp__claude-in-chrome__computer({
  action: "wait",
  duration: 2,
  tabId: tabId
});

// 5. Screenshot initial state
await mcp__claude-in-chrome__computer({
  action: "screenshot",
  tabId: tabId
});

Phase 3: Structure Analysis

Purpose: Understand page DOM structure before extraction.

Process:

1. Read page accessibility tree (read_page)
2. Identify data container elements
3. Map field selectors to schema
4. Verify selectors exist
5. Detect dynamic content patterns

Implementation:

// Read page structure
const pageTree = await mcp__claude-in-chrome__read_page({
  tabId: tabId,
  filter: "all"  // Get all elements for analysis
});

// For interactive elements only
const interactive = await mcp__claude-in-chrome__read_page({
  tabId: tabId,
  filter: "interactive"  // Buttons, links, inputs
});

// Find specific elements
const products = await mcp__claude-in-chrome__find({
  query: "product cards or listing items",
  tabId: tabId
});

const prices = await mcp__claude-in-chrome__find({
  query: "price elements",
  tabId: tabId
});

Phase 4: Data Extraction

Purpose: Extract data according to defined schema.

Process:

1. For simple text: Use get_page_text
2. For structured data: Use javascript_tool
3. For specific elements: Use find + read_page
4. Handle lazy-loaded content with scroll

Extraction Methods:

Method 1: Full Page Text (simplest, for articles)

const text = await mcp__claude-in-chrome__get_page_text({
  tabId: tabId
});

Method 2: JavaScript DOM Query (most flexible)

const data = await mcp__claude-in-chrome__javascript_tool({
  action: "javascript_exec",
  tabId: tabId,
  text: `
    // Extract product data
    Array.from(document.querySelectorAll('.product-card')).map(card => ({
      name: card.querySelector('.product-title')?.textContent?.trim(),
      price: parseFloat(card.querySelector('.price')?.textContent?.replace(/[^0-9.]/g, '')),
      url: card.querySelector('a')?.href,
      image: card.querySelector('img')?.src
    }))
  `
});

Method 3: Table Extraction

const tableData = await mcp__claude-in-chrome__javascript_tool({
  action: "javascript_exec",
  tabId: tabId,
  text: `
    // Extract table data
    const table = document.querySelector('table');
    const headers = Array.from(table.querySelectorAll('th')).map(th => th.textContent.trim());
    const rows = Array.from(table.querySelectorAll('tbody tr')).map(tr => {
      const cells = Array.from(tr.querySelectorAll('td')).map(td => td.textContent.trim());
      return headers.reduce((obj, header, i) => ({ ...obj, [header]: cells[i] }), {});
    });
    rows
  `
});

Method 4: Handling Lazy-Loaded Content

// Scroll to load content
await mcp__claude-in-chrome__computer({
  action: "scroll",
  scroll_direction: "down",
  scroll_amount: 5,
  tabId: tabId,
  coordinate: [500, 500]
});

// Wait for content to load
await mcp__claude-in-chrome__computer({
  action: "wait",
  duration: 2,
  tabId: tabId
});

// Now extract
const data = await extractData(tabId);

Phase 5: Pagination Handling

Purpose: Navigate through multiple pages and collect all data.

Pagination Patterns:

Pattern A: Numbered Pages

async function extractNumberedPages(baseUrl, maxPages, tabId) {
  const allData = [];

  for (let page = 1; page <= maxPages; page++) {
    const url = `${baseUrl}?page=${page}`;
    await mcp__claude-in-chrome__navigate({ url, tabId });
    await mcp__claude-in-chrome__computer({ action: "wait", duration: 2, tabId });

    const pageData = await extractPageData(tabId);
    if (pageData.length === 0) break;  // No more data

    allData.push(...pageData);

    // Rate limiting
    await mcp__claude-in-chrome__computer({ action: "wait", duration: 2, tabId });
  }

  return allData;
}

Pattern B: Next Button

async function extractWithNextButton(tabId) {
  const allData = [];
  let hasNext = true;

  while (hasNext) {
    const pageData = await extractPageData(tabId);
    allData.push(...pageData);

    // Find next button
    const nextButton = await mcp__claude-in-chrome__find({
      query: "next page button or pagination next link",
      tabId: tabId
    });

    if (nextButton && nextButton.elements && nextButton.elements.length > 0) {
      await mcp__claude-in-chrome__computer({
        action: "left_click",
        ref: nextButton.elements[0].ref,
        tabId: tabId
      });
      await mcp__claude-in-chrome__computer({ action: "wait", duration: 2, tabId });
    } else {
      hasNext = false;  // No more pages
    }
  }

  return allData;
}

Pattern C: Infinite Scroll

async function extractInfiniteScroll(tabId, maxScrolls = 20) {
  const allData = [];
  let previousCount = 0;
  let scrollCount = 0;

  while (scrollCount < maxScrolls) {
    const pageData = await extractPageData(tabId);

    if (pageData.length === previousCount) {
      // No new content loaded, done
      break;
    }

    allData.length = 0;
    allData.push(...pageData);
    previousCount = pageData.length;

    // Scroll down
    await mcp__claude-in-chrome__computer({
      action: "scroll",
      scroll_direction: "down",
      scroll_amount: 5,
      tabId: tabId,
      coordinate: [500, 500]
    });

    await mcp__claude-in-chrome__computer({ action: "wait", duration: 2, tabId });
    scrollCount++;
  }

  return allData;
}

Phase 6: Data Transformation

Purpose: Convert extracted data to requested output format.

JSON Output:

function toJSON(data, pretty = true) {
  return pretty ? JSON.stringify(data, null, 2) : JSON.stringify(data);
}

CSV Output:

function toCSV(data) {
  if (data.length === 0) return "";

  const headers = Object.keys(data[0]);
  const headerRow = headers.join(",");

  const dataRows = data.map(row =>
    headers.map(h => {
      const val = row[h];
      // Escape quotes and wrap in quotes if contains comma
      if (typeof val === "string" && (val.includes(",") || val.includes('"'))) {
        return `"${val.replace(/"/g, '""')}"`;
      }
      return val;
    }).join(",")
  );

  return [headerRow, ...dataRows].join("\n");
}

Markdown Table Output:

function toMarkdownTable(data) {
  if (data.length === 0) return "";

  const headers = Object.keys(data[0]);
  const headerRow = `| ${headers.join(" | ")} |`;
  const separator = `| ${headers.map(() => "---").join(" | ")} |`;

  const dataRows = data.map(row =>
    `| ${headers.map(h => String(row[h] || "")).join(" | ")} |`
  );

  return [headerRow, separator, ...dataRows].join("\n");
}

Phase 7: Storage

Purpose: Persist extracted data to Memory MCP for future use.

Implementation:

// Store in Memory MCP
await memory_store({
  namespace: `skills/tooling/web-scraping/${projectName}/${timestamp}`,
  data: {
    extraction_metadata: {
      source_url: targetUrl,
      extraction_date: new Date().toISOString(),
      record_count: data.length,
      output_format: outputFormat,
      pages_scraped: pageCount
    },
    extracted_data: data
  },
  tags: {
    WHO: `web-scraping-${sessionId}`,
    WHEN: new Date().toISOString(),
    PROJECT: projectName,
    WHY: "data-extraction",
    data_type: dataType,
    record_count: data.length
  }
});

LEARNED PATTERNS

This section is populated by Loop 1.5 (Session Reflection) as patterns are discovered.

High Confidence [conf:0.90]

No patterns captured yet. Patterns will be added as the skill is used and learnings are extracted.

Medium Confidence [conf:0.75]

No patterns captured yet.

Low Confidence [conf:0.55]

No patterns captured yet.

Success Criteria

Quality Thresholds:

• All targeted data fields extracted (100% schema coverage)
• Data validation passes for 95%+ of records
• Pagination handled completely (no missing pages)
• Output format matches requested format
• Rate limiting respected (no 429 errors)
• No page state modifications occurred
• Extraction completed within reasonable time (5 min per 100 records)

Failure Indicators:

• Schema validation fails for >5% of records
• Missing required fields in output
• Pagination incomplete (detected more pages than scraped)
• Rate limit errors encountered
• Page state modified (form submitted, button clicked for action)
• CAPTCHA or access denial encountered

MCP Integration

Required MCPs:

Optional MCPs:

• filesystem (for saving extracted data locally)

Memory Namespace

Pattern: skills/tooling/web-scraping/{project}/{timestamp}

Store:

• Extraction schemas (field definitions)
• Extracted datasets (structured data)
• Selector patterns (for similar pages)
• Error logs (selector failures, schema drift)

Retrieve:

• Similar extraction tasks (vector search by description)
• Proven selectors for known sites
• Historical extraction patterns

Tagging:

{
  "WHO": "web-scraping-{session_id}",
  "WHEN": "ISO8601_timestamp",
  "PROJECT": "{project_name}",
  "WHY": "data-extraction",
  "source_domain": "example.com",
  "data_type": "products|articles|listings|tables",
  "record_count": 150,
  "pages_scraped": 5
}

Examples

Example 1: Product Catalog Extraction

Complexity: Medium (structured data, pagination)

Task: Extract product listings from e-commerce category page

Planning Output (sequential-thinking):

Thought 1/6: Need to extract product name, price, URL, and image from catalog
Thought 2/6: Schema: {name: string, price: number, url: url, image: url}
Thought 3/6: Identify product card containers via read_page
Thought 4/6: Use javascript_tool to query all product cards
Thought 5/6: Detect pagination pattern (numbered pages with ?page=N)
Thought 6/6: Output format: JSON array, validate price > 0

Execution:

// 1. Navigate to catalog
await navigate({ url: "https://example.com/products", tabId });

// 2. Analyze structure
const structure = await read_page({ tabId, filter: "all" });

// 3. Extract data
const products = await javascript_tool({
  action: "javascript_exec",
  tabId,
  text: `
    Array.from(document.querySelectorAll('.product-card')).map(card => ({
      name: card.querySelector('.title')?.textContent?.trim(),
      price: parseFloat(card.querySelector('.price')?.textContent?.replace(/[^0-9.]/g, '')),
      url: card.querySelector('a')?.href,
      image: card.querySelector('img')?.src
    }))
  `
});

// 4. Handle pagination (repeat for each page)
// 5. Transform to JSON
const output = JSON.stringify(products, null, 2);

Result: 150 products extracted across 5 pages

Output Format: JSON

[
  {
    "name": "Product A",
    "price": 29.99,
    "url": "https://example.com/products/a",
    "image": "https://example.com/images/a.jpg"
  }
]

Example 2: Article Content Scraping

Complexity: Simple (single page, text focus)

Task: Extract article headline, author, date, and body text

Planning Output (sequential-thinking):

Thought 1/4: Single page extraction, no pagination needed
Thought 2/4: Schema: {headline: string, author: string, date: date, body: string}
Thought 3/4: Use get_page_text for full content, javascript_tool for metadata
Thought 4/4: Output format: Markdown for readability

Execution:

// 1. Navigate to article
await navigate({ url: "https://news.example.com/article/123", tabId });

// 2. Extract metadata
const metadata = await javascript_tool({
  action: "javascript_exec",
  tabId,
  text: `({
    headline: document.querySelector('h1')?.textContent?.trim(),
    author: document.querySelector('.author')?.textContent?.trim(),
    date: document.querySelector('time')?.getAttribute('datetime')
  })`
});

// 3. Extract body text
const body = await get_page_text({ tabId });

// 4. Combine
const article = { ...metadata, body };

Result: Article content extracted

Output Format: Markdown

# Article Headline

**Author**: Jane Doe
**Date**: 2026-01-12

Article body text here...

Example 3: Table Data Extraction

Complexity: Simple (structured HTML table)

Task: Extract financial data from HTML table

Planning Output (sequential-thinking):

Thought 1/4: HTML table with headers in <th>, data in <td>
Thought 2/4: Use javascript_tool to parse table structure
Thought 3/4: Convert to array of objects with header keys
Thought 4/4: Output format: CSV for spreadsheet import

Execution:

// 1. Navigate to page
await navigate({ url: "https://finance.example.com/data", tabId });

// 2. Extract table
const tableData = await javascript_tool({
  action: "javascript_exec",
  tabId,
  text: `
    const table = document.querySelector('table.financial-data');
    const headers = Array.from(table.querySelectorAll('th')).map(th => th.textContent.trim());
    Array.from(table.querySelectorAll('tbody tr')).map(tr => {
      const cells = Array.from(tr.querySelectorAll('td')).map(td => td.textContent.trim());
      return headers.reduce((obj, header, i) => ({ ...obj, [header]: cells[i] }), {});
    })
  `
});

// 3. Transform to CSV
const csv = toCSV(tableData);

Result: 50 rows extracted

Output Format: CSV

Date,Open,High,Low,Close,Volume
2026-01-12,150.00,152.50,149.00,151.75,1000000
2026-01-11,148.50,150.25,147.00,150.00,950000

Example 4: Multi-Page Directory Scraping

Complexity: High (pagination, rate limiting)

Task: Extract business listings from directory with “Load More” pagination

Planning Output (sequential-thinking):

Thought 1/8: Directory with business name, phone, address, website
Thought 2/8: Schema: {name: string, phone: string, address: string, website: url}
Thought 3/8: Pagination: "Load More" button at bottom
Thought 4/8: Rate limit: 3 seconds between loads
Thought 5/8: Max iterations: 20 (to prevent infinite loops)
Thought 6/8: End condition: "Load More" button disappears or disabled
Thought 7/8: Output format: JSON
Thought 8/8: Store in Memory MCP for analysis

Result: 200 businesses extracted across 20 “Load More” clicks

Execution Time: 90 seconds (with rate limiting)

Anti-Patterns to Avoid

Related Skills

Upstream (provide input to this skill):

• intent-analyzer – Detect web scraping requirements
• prompt-architect – Optimize extraction descriptions
• planner – High-level data collection strategy

Downstream (use output from this skill):

• data-analysis – Analyze extracted datasets
• reporting – Generate reports from data
• ml-pipeline – Use data for training

Parallel (work together):

• browser-automation – For interactive pre-requisites (login)
• api-integration – Hybrid scraping/API workflows

Selector Stability Strategies

Problem: Web pages change, selectors break.

Strategies:

Best Practices:

1. Prefer data-* attributes (designed for scripting)
2. Use semantic class names over positional selectors
3. Combine multiple attributes for specificity
4. Use find tool with natural language as fallback
5. Log selector failures to detect schema drift

Output Format Templates

JSON Template

{
  "extraction_metadata": {
    "source_url": "https://example.com/products",
    "extraction_date": "2026-01-12T10:30:00Z",
    "record_count": 150,
    "pages_scraped": 5
  },
  "data": [
    {
      "name": "Product Name",
      "price": 29.99,
      "url": "https://example.com/product/1",
      "image": "https://example.com/images/1.jpg"
    }
  ]
}

CSV Template

name,price,url,image
"Product Name",29.99,"https://example.com/product/1","https://example.com/images/1.jpg"

Markdown Table Template

| Name | Price | URL |
|------|-------|-----|
| Product Name | $29.99 | [Link](https://example.com/product/1) |

Maintenance & Updates

Version History:

• v1.0.0 (2026-01-12): Initial release with READ-only focus, pagination handling, rate limiting

Feedback Loop:

• Loop 1.5 (Session): Store learnings from extractions
• Loop 3 (Meta-Loop): Aggregate patterns every 3 days
• Update LEARNED PATTERNS section with new discoveries

Continuous Improvement:

• Monitor extraction success rate via Memory MCP
• Identify common selector failures for pattern updates
• Optimize extraction strategies based on site patterns

WEB_SCRAPING_VERILINGUA_VERIX_COMPLIANT