name: visual-browser
description: Browser automation using Computer Vision (Gemini 3 Flash). Use as fallback when agent-browser fails on icon-only buttons, complex Canvas/JS UIs, or visual audits.
user_invocable: false

Visual Browser Skill

This skill allows the agent to interact with the web using Computer Vision and spatial reasoning via the browser-use library and Gemini 3 Flash. Unlike text-only scrapers, it can "see" the page, handle complex layouts, and interact with elements based on their visual appearance.

Usage

Invoke this skill when you need to perform a task that requires visual understanding of a website (e.g., clicking complex buttons, navigating non-standard UIs, or summarizing visual content).

uv run /Users/terry/skills/visual-browser/browse.py "Find the latest AI news on TechCrunch and summarize the top 3 stories."

Capabilities

• Visual Navigation: Identifies elements by their visual features, not just HTML selectors.
• Dynamic Interaction: Handles JS-heavy sites, modals, and complex state changes.
• Multimodal Reasoning: Uses Gemini 2.0 Flash's vision capabilities to plan and execute browser actions.

Configuration

• Requires GOOGLE_API_KEY in the environment.
• Uses playwright for browser control.
• Automatically installs dependencies via uv run.

Best Practices

• Use for UI-heavy sites where standard scrapers fail.
• Provide clear, step-by-step or outcome-oriented tasks.
• For simple text retrieval, consider WebSearch or WebFetch instead to save tokens.