browser

Browser Automation Skill

This skill handles browser automation tasks using subagent architecture to keep the main context window clean.

Core Principle

NEVER call Claude-in-Chrome MCP tools directly in main context.

All browser automation is delegated to the browser-automation Task subagent which returns concise summaries.

Detection Triggers

Activate this skill when detecting:

Web Testing

• "test the website", "check the page", "verify the UI"
• "screenshot", "capture the screen", "take a picture"
• "does the site work", "is the page loading"

Form Automation

• "fill out the form", "submit this form", "enter data"
• "sign up", "register", "create account" (with explicit user permission)
• "login", "authenticate" (without entering passwords directly)

Data Extraction

• "scrape this page", "extract data from", "get the text"
• "what's on the page", "read the content", "parse the data"
• "find the price", "get the title", "extract the links"

Workflow Automation

• "automate this workflow", "record these steps", "create a GIF"
• "click through", "navigate to", "go to page"
• "test the flow", "verify the process"

Debugging & Monitoring

• "check console logs", "monitor network requests", "debug the page"
• "why is this failing", "what's the error", "inspect the page"

Subagent Delegation Pattern

When a browser automation task is detected:

Task("Browser: {brief_description}

**Task:** {user_request}
**URL:** {target_url if provided}
**Actions:** {list of expected actions}

Using Claude-in-Chrome MCP tools:
1. Start with tabs_context_mcp to get current tab state
2. Create new tab or use existing based on context
3. Navigate, interact, extract data as needed
4. Capture screenshots/GIFs for multi-step workflows
5. Monitor console/network if debugging

Return ONLY a concise summary:
- Actions taken (e.g., "Navigated to X, filled form, clicked submit")
- Key findings (e.g., "Form submitted successfully", "Error: XYZ")
- Screenshots/visual evidence if captured
- Any errors or unexpected behavior

Do NOT include:
- Raw HTML or page source
- Verbose console logs (use pattern filtering)
- Detailed network request payloads
- Implementation details

subagent_type: browser-automation
")