---

name: browser-research description: | This skill should be used when the user needs to perform deep web research across multiple websites, extract data from JavaScript-rendered or dynamic pages, navigate through search results or paginated content, or build comprehensive research reports from multiple sources.

Use this skill when the user says: "research this topic across sites", "scrape this website", "extract data from this page", "compare prices", "browse and extract information", "fill forms and get results", "navigate through pages", "click through search results", "research authenticated content", or "cross-reference multiple sources".

Prefer this over WebFetch when pages require JavaScript, interaction, authentication, or multi-step navigation. allowed-tools: Bash(agent-browser:*)

Browser Research

Execute deep web research missions using agent-browser for interactive page navigation and data extraction.

When to Use

• Multi-page research: Navigating search results, following links, exploring sites
• Dynamic content: JavaScript-rendered pages, SPAs, interactive dashboards
• Form interaction: Filling search fields, applying filters, submitting queries
• Data extraction: Tables, lists, product information, structured data
• Authenticated content: Sites requiring login or session state

For simple static pages, consider WebFetch first. Escalate to agent-browser when content is incomplete or interaction is required.

Core Research Loop

Copy┌─────────────────────────────────────────────────────┐
│  1. OPEN     agent-browser open <url>               │
│  2. SNAPSHOT agent-browser snapshot -i              │
│  3. EXTRACT  agent-browser get text @ref            │
│  4. NAVIGATE agent-browser click @ref / fill / etc  │
│  5. REPEAT   Loop steps 2-4 until mission complete  │
└─────────────────────────────────────────────────────┘

The Snapshot-Act Pattern

Always snapshot before interacting. Refs (@e1, @e2) are stable identifiers for elements:

Copyagent-browser open "https://example.com/search"
agent-browser snapshot -i                    # See what's available
agent-browser fill @e3 "search query"        # Fill the search box
agent-browser click @e5                      # Click search button
agent-browser wait --load networkidle        # Wait for results
agent-browser snapshot -i                    # See new state

Research Planning

Before starting, define:

Source Discovery

Start broad, then narrow:

Copy# Option 1: Direct navigation (when you know the site)
agent-browser open "https://known-site.com/pricing"

# Option 2: Search engine discovery
agent-browser open "https://www.google.com"
agent-browser snapshot -i
agent-browser fill @e1 "CRM pricing comparison 2025"
agent-browser press Enter
agent-browser wait --load networkidle
agent-browser snapshot -i
# Extract relevant URLs from results

Scenario Recipes

Use recipes when you need a repeatable pattern. Each recipe should include:

1. Intent - The question being answered and the expected output shape
2. Minimal Steps - 3-6 steps with snapshots, verification, and stop conditions
3. Output - JSON fields with sources, access dates, and missing-data notes

Extraction

• Data Extraction Patterns - Feed top-N, paginated catalog, filtered results

Comparison

• Multi-Source Research - Consensus checks, price/spec comparisons, recency validation

Exploration

• Site Exploration - Docs crawl, category discovery, auth-gated inventory

Data Extraction Strategies

Extract Text Content

Copy# Single element
agent-browser get text @e15

# Multiple elements - snapshot shows structure
agent-browser snapshot -s ".results"         # Scope to results section
agent-browser get text @e20                  # First result
agent-browser get text @e25                  # Second result

Handle Pagination

Copy# Pattern: Extract current page, navigate to next, repeat
agent-browser snapshot -i
# ... extract data from current page ...
agent-browser click @e99                     # "Next" button
agent-browser wait --load networkidle
agent-browser snapshot -i                    # New page content

Handle Lazy Loading

Copyagent-browser scroll down 1000
agent-browser wait 1000                      # Allow content to load
agent-browser snapshot -i                    # See newly loaded content

Output Format

Structure your research results for clarity:

Copy{
  "mission": "Original research question",
  "sources": [
    {
      "url": "https://example.com/page",
      "title": "Page Title",
      "accessed": "2025-01-19",
      "data": {
        "key_finding_1": "value",
        "key_finding_2": "value"
      }
    }
  ],
  "summary": "Key findings in 2-3 sentences",
  "limitations": "Any access issues or incomplete data"
}

Always include:

• Source URLs for attribution
• Access timestamps for currency
• Limitations for transparency

Sessions for Parallel Research

Use named sessions to research multiple sources simultaneously:

Copy# Research source A
agent-browser --session src-a open "https://site-a.com"
agent-browser --session src-a snapshot -i

# Research source B (parallel)
agent-browser --session src-b open "https://site-b.com"
agent-browser --session src-b snapshot -i

# Continue working in each session
agent-browser --session src-a click @e5
agent-browser --session src-b fill @e3 "query"

Quick Reference

For complete command documentation, see: agent-browser skill

See Also

• Data Extraction Patterns - Tables, lists, structured data
• Multi-Source Research - Comparing across sites
• Site Exploration - Deep-diving single domains

Handling Failures

Research automation can fail. Handle common issues:

Copy# Check if element exists before interacting
agent-browser is visible @e50
# If false, the page structure may have changed - re-snapshot

# Timeout waiting for content
agent-browser wait --load networkidle
# If this times out, content may require scrolling or interaction

# Element not found - try semantic locators
agent-browser find text "Submit" click
# More resilient than refs when DOM structure changes

# Network errors - check page loaded
agent-browser get url
# Verify you're on the expected page before extracting

When failures occur:

• Re-snapshot - DOM may have changed
• Try find locators - More stable than refs for dynamic content
• Check URL - Redirects or errors may have changed the page
• Report limitations - Note access issues in output, don't silently skip

Best Practices

1. Snapshot frequently - DOM changes after interactions
2. Use -i flag - Interactive elements only, cleaner output
3. Wait after navigation - --load networkidle ensures content is ready
4. Scope when needed - -s ".container" focuses on relevant sections
5. Track sources - Use get url and get title for attribution
6. Handle failures - Report access issues, don't silently skip
7. Use find for dynamic content - When refs change, semantic locators are more stable
8. Close when done - agent-browser close releases resources