web-scraper（网络爬虫）

Web Scraper

Overview

Web scraping inteligente multi-estrategia. Extrai dados estruturados de paginas web (tabelas, listas, precos). Paginacao, monitoramento e export CSV/JSON.

When to Use This Skill

• When the user mentions "scraper" or related topics
• When the user mentions "scraping" or related topics
• When the user mentions "extrair dados web" or related topics
• When the user mentions "web scraping" or related topics
• When the user mentions "raspar dados" or related topics
• When the user mentions "coletar dados site" or related topics

Do Not Use This Skill When

• The task is unrelated to web scraper
• A simpler, more specific tool can handle the request
• The user needs general-purpose assistance without domain expertise

How It Works

Execute phases in strict order. Each phase feeds the next.

1. CLARIFY  ->  2. RECON  ->  3. STRATEGY  ->  4. EXTRACT  ->  5. TRANSFORM  ->  6. VALIDATE  ->  7. FORMAT

Never skip Phase 1 or Phase 2. They prevent wasted effort and failed extractions.

Fast path: If user provides URL + clear data target + the request is simple (single page, one data type), compress Phases 1-3 into a single action: fetch, classify, and extract in one WebFetch call. Still validate and format.

Capabilities

• Multi-strategy: WebFetch (static), Browser automation (JS-rendered), Bash/curl (APIs), WebSearch (discovery)
• Extraction modes: table, list, article, product, contact, FAQ, pricing, events, jobs, custom
• Output formats: Markdown tables (default), JSON, CSV
• Pagination: auto-detect and follow (page numbers, infinite scroll, load-more)
• Multi-URL: extract same structure across sources with comparison and diff
• Validation: confidence ratings (HIGH/MEDIUM/LOW) on every extraction
• Auto-escalation: WebFetch fails silently -> automatic Browser fallback
• Data transforms: cleaning, normalization, deduplication, enrichment
• Differential mode: detect changes between scraping runs

Web Scraper

Multi-strategy web data extraction with intelligent approach selection, automatic fallback escalation, data transformation, and structured output.

Phase 1: Clarify

Establish extraction parameters before touching any URL.

Required Parameters

Optional Parameters

Clarification Rules

• If user provides a URL and clear data target, proceed directly to Phase 2. Do NOT ask unnecessary questions.
• If request is ambiguous (e.g. "scrape this site"), ask ONLY: "What specific data do you want me to extract from this page?"
• Default to Markdown table output. Mention alternatives only if relevant.
• Accept requests in any language. Always respond in the user's language.
• If user says "everything" or "all data", perform recon first, then present what's available and let user choose.

Discovery Mode

When user has a topic but no specific URL:

1. Use WebSearch to find the most relevant pages
2. Present top 3-5 URLs with descriptions
3. Let user choose which to scrape, or scrape all
4. Proceed to Phase 2 with selected URL(s)

Example: "find and extract pricing data for CRM tools" -> WebSearch("CRM tools pricing comparison 2026") -> Present top results -> User selects -> Extract

Phase 2: Reconnaissance

Analyze the target page before extraction.

Step 2.1: Initial Fetch

Use WebFetch to retrieve and analyze the page structure:

WebFetch(
  url = TARGET_URL,
  prompt = "Analyze this page structure and report:
    1. Page type: article, product listing, search results, data table,
       directory, dashboard, API docs, FAQ, pricing page, job board, events, or other
    2. Main content structure: tables, ordered/unordered lists, card grid, free-form text,
       accordion/collapsible sections, tabs
    3. Approximate number of distinct data items visible
    4. JavaScript rendering indicators: empty containers, loading spinners,
       SPA framework markers (React root, Vue app, Angular), minimal HTML with heavy JS
    5. Pagination: next/prev links, page numbers, load-more buttons,
       infinite scroll indicators, total results count
    6. Data density: how much structured, extractable data exists
    7. List the main data fields/columns available for extraction
    8. Embedded structured data: JSON-LD, microdata, OpenGraph tags
    9. Available download links: CSV, Excel, PDF, API endpoints"
)

Step 2.2: Evaluate Fetch Quality

Step 2.3: Content Classification

Classify into an extraction mode:

Record: page type, extraction mode, JS rendering needed (yes/no), available fields, structured data present (JSON-LD etc.).

If user asked for "everything", present the available fields and let them choose.

Phase 3: Strategy Selection

Choose the extraction approach based on recon results.

Decision Tree

Structured data (JSON-LD, microdata) has what we need?
 |
 +-- YES --> STRATEGY E: Extract structured data directly
 |
 +-- NO: Content fully visible in WebFetch?
      |
      +-- YES: Need precise element targeting?
      |    |
      |    +-- NO  --> STRATEGY A: WebFetch + AI extraction
      |    +-- YES --> STRATEGY B: Browser automation
      |
      +-- NO: JavaScript rendering detected?
           |
           +-- YES --> STRATEGY B: Browser automation
           +-- NO:  API/JSON/XML endpoint or download link?
                |
                +-- YES --> STRATEGY C: Bash (curl + jq)
                +-- NO  --> Report access issue to user

Strategy A: Webfetch With Ai Extraction

Best for: Static pages, articles, simple tables, well-structured HTML.

Use WebFetch with a targeted extraction prompt tailored to the mode:

WebFetch(
  url = URL,
  prompt = "Extract [DATA_TARGET] from this page.
    Return ONLY the extracted data as [FORMAT] with these columns/fields: [FIELDS].
    Rules:
    - If a value is missing or unclear, use 'N/A'
    - Do not include navigation, ads, footers, or unrelated content
    - Preserve original values exactly (numbers, currencies, dates)
    - Include ALL matching items, not just the first few
    - For each item, also extract the URL/link if available"
)

Auto-escalation: If WebFetch returns suspiciously few items (less than 50% of expected from recon), or mostly empty fields, automatically escalate to Strategy B without asking user. Log the escalation in notes.

Strategy B: Browser Automation

Best for: JS-rendered pages, SPAs, interactive content, lazy-loaded data.

Sequence:

1. Get tab context: tabs_context_mcp(createIfEmpty=true) -> get tabId
2. Navigate to URL: navigate(url=TARGET_URL, tabId=TAB)
3. Wait for content to load: computer(action="wait", duration=3, tabId=TAB)
4. Check for cookie/consent banners: find(query="cookie consent or accept button", tabId=TAB)

- If found, dismiss it (prefer privacy-preserving option)

1. Read page structure: read_page(tabId=TAB) or get_page_text(tabId=TAB)
2. Locate target elements: find(query="[DESCRIPTION]", tabId=TAB)
3. Extract with JavaScript for precise data via javascript_tool

// Table extraction
const rows = document.querySelectorAll('TABLE_SELECTOR tr');
const data = Array.from(rows).map(row => {
  const cells = row.querySelectorAll('td, th');
  return Array.from(cells).map(c => c.textContent.trim());
});
JSON.stringify(data);

// List/card extraction
const items = document.querySelectorAll('ITEM_SELECTOR');
const data = Array.from(items).map(item => ({
  field1: item.querySelector('FIELD1_SELECTOR')?.textContent?.trim() || null,
  field2: item.querySelector('FIELD2_SELECTOR')?.textContent?.trim() || null,
  link: item.querySelector('a')?.href || null,
}));
JSON.stringify(data);

1. For lazy-loaded content, scroll and re-extract: computer(action="scroll", scroll_direction="down", tabId=TAB) then computer(action="wait", duration=2, tabId=TAB)

Strategy C: Bash (Curl + Jq)

Best for: REST APIs, JSON endpoints, XML feeds, CSV/Excel downloads.

## Json Api

curl -s "API_URL" | jq '[.items[] | {field1: .key1, field2: .key2}]'

## Csv Download

curl -s "CSV_URL" -o /tmp/scraped_data.csv

## Xml Parsing

curl -s "XML_URL" | python3 -c "
import xml.etree.ElementTree as ET, json, sys
tree = ET.parse(sys.stdin)

## ... Parse And Output Json

"

Strategy D: Hybrid

When a single strategy is insufficient, combine:

1. WebSearch to discover relevant URLs
2. WebFetch for initial content assessment
3. Browser automation for JS-heavy sections
4. Bash for post-processing (jq, python for data cleaning)

Strategy E: Structured Data Extraction

When JSON-LD, microdata, or OpenGraph is present:

1. Use Browser javascript_tool to extract structured data:

const scripts = document.querySelectorAll('script[type="application/ld+json"]');
const data = Array.from(scripts).map(s => {
  try { return JSON.parse(s.textContent); } catch { return null; }
}).filter(Boolean);
JSON.stringify(data);

1. This often provides cleaner, more reliable data than DOM scraping
2. Fall back to DOM extraction only for fields not in structured data

Pagination Handling

When pagination is detected and user wants multiple pages:

Page-number pagination (any strategy):

1. Extract data from current page
2. Identify URL pattern (e.g. ?page=N, /page/N, &offset=N)
3. Iterate through pages up to user's max (default: 5 pages)
4. Show progress: "Extracting page 2/5..."
5. Concatenate all results, deduplicate if needed

Infinite scroll (Browser only):

1. Extract currently visible data
2. Record item count
3. Scroll down: computer(action="scroll", scroll_direction="down", tabId=TAB)
4. Wait: computer(action="wait", duration=2, tabId=TAB)
5. Extract newly loaded data
6. Compare count - if no new items after 2 scrolls, stop
7. Repeat until no new content or max iterations (default: 5)

"Load More" button (Browser only):

1. Extract currently visible data
2. Find button: find(query="load more button", tabId=TAB)
3. Click it: computer(action="left_click", ref=REF, tabId=TAB)
4. Wait and extract new content
5. Repeat until button disappears or max iterations reached

Phase 4: Extract

Execute the selected strategy using mode-specific patterns. See references/extraction-patterns.md for CSS selectors and JavaScript snippets.

Table Mode

WebFetch prompt:

"Extract ALL rows from the table(s) on this page.
Return as a markdown table with exact column headers.
Include every row - do not truncate or summarize.
Preserve numeric precision, currencies, and units."

List Mode

WebFetch prompt:

"Extract each [ITEM_TYPE] from this page.
For each item, extract: [FIELD_LIST].
Return as a JSON array of objects with these keys: [KEY_LIST].
Include ALL items, not just the first few. Include link/URL for each item if available."

Article Mode

WebFetch prompt:

"Extract article metadata:
- title, author, date, tags/categories, word count estimate
- Key factual data points, statistics, and named entities
Return as structured markdown. Summarize the content; do not reproduce full text."

Product Mode

WebFetch prompt:

"Extract product data with these exact fields:
- name, brand, price, currency, originalPrice (if discounted),
  availability, description (first 200 chars), rating, reviewCount,
  specifications (as key-value pairs), productUrl, imageUrl
Return as JSON. Use null for missing fields."

Also check for JSON-LD Product schema (Strategy E) first.

Contact Mode

WebFetch prompt:

"Extract contact information for each person/entity:
- name, title, role, email, phone, address, organization, website, linkedinUrl
Return as a markdown table. Only extract real contacts visible on the page."

Faq Mode

WebFetch prompt:

"Extract all question-answer pairs from this page.
For each FAQ item extract:
- question: the exact question text
- answer: the answer text (first 300 chars if long)
- category: the section/category if grouped
Return as a JSON array of objects."

Pricing Mode

WebFetch prompt:

"Extract all pricing plans/tiers from this page.
For each plan extract:
- planName, monthlyPrice, annualPrice, currency
- features (array of included features)
- limitations (array of limits or excluded features)
- ctaText (call-to-action button text)
- highlighted (true if marked as recommended/popular)
Return as JSON. Use null for missing fields."

Events Mode

WebFetch prompt:

"Extract all events/sessions from this page.
For each event extract:
- title, date, time, endTime, location, description (first 200 chars)
- speakers (array of names), category, registrationUrl
Return as JSON. Use null for missing fields."

Jobs Mode

WebFetch prompt:

"Extract all job listings from this page.
For each job extract:
- title, company, location, salary, salaryRange, type (full-time/part-time/contract)
- postedDate, description (first 200 chars), applyUrl, tags
Return as JSON. Use null for missing fields."

Custom Mode

When user provides specific selectors or field descriptions:

• Use Browser automation with javascript_tool and user's CSS selectors
• Or use WebFetch with a prompt built from user's field descriptions
• Always confirm extracted schema with user before proceeding to multi-URL

Multi-Url Extraction

When extracting from multiple URLs:

1. Extract from the first URL to establish the data schema
2. Show user the first results and confirm the schema is correct
3. Extract from remaining URLs using the same schema
4. Add a source column/field to every record with the origin URL
5. Combine all results into a single output
6. Show progress: "Extracting 3/7 URLs..."

Phase 5: Transform

Clean, normalize, and enrich extracted data before validation. See references/data-transforms.md for patterns.

Automatic Transforms (Always Apply)

Conditional Transforms (Apply When Relevant)

Data Enrichment (Only When Useful)

Deduplication Strategy

When combining data from multiple pages or URLs:

1. Exact match: rows with identical values in all fields -> keep first
2. Near match: rows with same key fields (name+source) but different details -> keep most complete (fewer nulls), flag in notes
3. Report: "Removed N duplicate rows" in delivery notes

Phase 6: Validate

Verify extraction quality before delivering results.

Validation Checks

Confidence Rating

Assign to every extraction:

Always report confidence with specifics:

Confidence: HIGH - 47 items extracted, all 6 fields populated, matches expected count from page analysis.

Auto-Recovery (Try Before Reporting Issues)

Log all recovery attempts in delivery notes. Inform user of any irrecoverable gaps with specific details.

Phase 7: Format And Deliver

Structure results according to user preference. See references/output-templates.md for complete formatting templates.

Delivery Envelope

ALWAYS wrap results with this metadata header:

## Extraction Results

**Source:** [Page Title](http://example.com)
**Date:** YYYY-MM-DD HH:MM UTC
**Items:** N records (M fields each)
**Confidence:** HIGH | MEDIUM | LOW
**Strategy:** A (WebFetch) | B (Browser) | C (API) | E (Structured Data)
**Format:** Markdown Table | JSON | CSV

---

[DATA HERE]

---

**Notes:**
- [Any gaps, issues, or observations]
- [Transforms applied: deduplication, normalization, etc.]
- [Pages scraped if paginated: "Pages 1-5 of 12"]
- [Auto-escalation if it occurred: "Escalated from WebFetch to Browser"]

Markdown Table Rules

• Left-align text columns (:---), right-align numbers (---:)
• Consistent column widths for readability
• Include summary row for numeric data when useful (totals, averages)
• Maximum 10 columns per table; split wider data into multiple tables or suggest JSON format
• Truncate long cell values to 60 chars with ... indicator
• Use N/A for missing values, never leave cells empty
• For multi-page results, show combined table (not per-page)

Json Rules

• Use camelCase for keys (e.g. productName, unitPrice)
• Wrap in metadata envelope: {"metadata": {"source": "URL", "title": "Page Title", "extractedAt": "ISO-8601", "itemCount": 47, "fieldCount": 6, "confidence": "HIGH", "strategy": "A", "transforms": ["deduplication", "priceNormalization"], "notes": []}, "data": [...]}
• Pretty-print with 2-space indentation
• Numbers as numbers (not strings), booleans as booleans
• null for missing values (not empty strings)

Csv Rules

• First row is always headers
• Quote any field containing commas, quotes, or newlines
• UTF-8 encoding with BOM for Excel compatibility
• Use , as delimiter (standard)
• Include metadata as comments: # Source: URL

File Output

When user requests file save:

• Markdown: .md extension
• JSON: .json extension
• CSV: .csv extension
• Confirm path before writing
• Report full file path and item count after saving

Multi-Url Comparison Format

When comparing data across multiple sources:

• Add Source as the first column/field
• Use short identifiers for sources (domain name or user label)
• Group by source or interleave based on user preference
• Highlight differences if user asks for comparison
• Include summary: "Best price: $X at store-b.com"

Differential Output

When user requests change detection (diff mode):

• Compare current extraction with previous run
• Mark new items with [NEW]
• Mark removed items with [REMOVED]
• Mark changed values with [WAS: old_value]
• Include summary: "Changes since last run: +5 new, -2 removed, 3 modified"

Rate Limiting

• Maximum 1 request per 2 seconds for sequential page fetches
• For multi-URL jobs, process sequentially with pauses
• If a site returns 429 (Too Many Requests), stop and report to user

Access Respect

• If a page blocks access (403, CAPTCHA, login wall), report to user
• Do NOT attempt to bypass bot detection, CAPTCHAs, or access controls
• Do NOT scrape behind authentication unless user explicitly provides access
• Respect robots.txt directives when known

Copyright

• Do NOT reproduce large blocks of copyrighted article text
• For articles: extract factual data, statistics, and structured info; summarize narrative content
• Always include source attribution (http://example.com) in output

Data Scope

• Extract ONLY what the user explicitly requested
• Warn user before collecting potentially sensitive data at scale (emails, phone numbers, personal information)
• Do not store or transmit extracted data beyond what the user sees

Failure Protocol

When extraction fails or is blocked:

1. Explain the specific reason (JS rendering, bot detection, login, etc.)
2. Suggest alternatives (different URL, API if available, manual approach)
3. Never retry aggressively or escalate access attempts

Quick Reference: Mode Cheat Sheet

References

• Extraction patterns: references/extraction-patterns.md CSS selectors, JavaScript snippets, JSON-LD parsing, domain tips.
• Output templates: references/output-templates.md Markdown, JSON, CSV templates with complete examples.
• Data transforms: references/data-transforms.md Cleaning, normalization, deduplication, enrichment patterns.

Best Practices

• Provide clear, specific context about your project and requirements
• Review all suggestions before applying them to production code
• Combine with other complementary skills for comprehensive analysis

Common Pitfalls

• Using this skill for tasks outside its domain expertise
• Applying recommendations without understanding your specific context
• Not providing enough project context for accurate analysis

Limitations

• Use this skill only when the task clearly matches the scope described above.
• Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
• Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.