The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides patterns for processing natural language commands to control web page interfaces, creating a surface for indirect prompt injection.
Ingestion points: The agent.execute(command) method is shown processing input from user-facing contexts, such as the SaaSCopilot.handleUserCommand and AccessibilityAgent.handleVoiceCommand examples in SKILL.md.
Capability inventory: The agent has capabilities to click elements, fill forms, navigate pages, and execute multi-step workflows using DOM manipulation.
Boundary markers: The provided code snippets do not include explicit delimiters or system-level instructions to ignore malicious commands embedded within the input data.
Sanitization: While the documentation mentions sanitizing user input as a best practice, no specific sanitization logic is implemented in the integration examples.
[EXTERNAL_DOWNLOADS]: The skill references and fetches assets from well-known external sources.
Downloads the Page Agent library via the JSDelivr CDN (https://cdn.jsdelivr.net/npm/page-agent@1.8.2/dist/iife/page-agent.demo.js).
Directs users to the official project repository and documentation hosted on Alibaba's GitHub organization.
[COMMAND_EXECUTION]: Instructions are provided for executing the Model Context Protocol (MCP) server.
Uses npx page-agent-mcp to launch the server, enabling external control of the browser automation agent.

page-agent-web-automation