The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill interacts with external, potentially untrusted websites, creating a surface for indirect prompt injection where malicious content on a page could attempt to influence the agent's behavior.
Ingestion points: The skill reads page content, console logs, and error messages from the target URL using the agent-browser tool (documented in SKILL.md).
Boundary markers: The instructions do not define specific delimiters or instructions to ignore embedded commands within the site content.
Capability inventory: The agent has the ability to write files (screenshots, videos, reports) and perform network interactions via the agent-browser tool.
Sanitization: No explicit sanitization or filtering of the content retrieved from the target website is mentioned.
[COMMAND_EXECUTION]: The skill instructions include the execution of shell commands to manage local files and orchestrate browser sessions.
Evidence: SKILL.md contains commands for directory creation (mkdir -p), file copying (cp), and browser control through agent-browser subcommands.

dogfood