The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill instructs the agent to use the eval command to execute JavaScript within the browser context for tasks like polling DOM values and determining if a page is fully hydrated. This is a standard feature of the 'agent-browser' tool used for complex state verification.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from the web, creating a potential surface for indirect prompt injection.\n
Ingestion points: Untrusted data enters the context through agent-browser open, snapshot -i, and get title commands as described in SKILL.md and references/command-cookbook.md.\n
Boundary markers: The skill does not explicitly instruct the agent to use delimiters or ignore instructions found within the processed web content.\n
Capability inventory: The skill allows for browser-side code execution (eval), file system writes (screenshots), and network navigation across all provided scripts.\n
Sanitization: No explicit sanitization or filtering of external content is mentioned, although the skill does include specific guidance on escaping special characters in PowerShell to prevent shell-level command confusion.\n- [COMMAND_EXECUTION]: The skill provides numerous recipes for executing shell commands. It specifically addresses risks associated with PowerShell's handling of special characters (e.g., the @ symbol in element references and redirection operators in complex strings) and provides instructions on how to avoid unintended file creation and command misinterpretation. The skill also discusses using the --no-sandbox flag to resolve browser startup failures on Windows, which reduces browser isolation but is a common practice in automated environments.\n- [EXTERNAL_DOWNLOADS]: The skill recommends installing the core tool via npm i -g agent-browser. This is a standard installation procedure using a well-known package registry.

use-agent-browser