The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted video titles and descriptions from YouTube. Maliciously crafted metadata could attempt to influence the agent's behavior. This is mitigated by explicit instructions in SKILL.md telling the agent to treat this content as data rather than instructions.
Ingestion points: Video metadata is fetched from YouTube RSS and the YouTube Data API in collect.py and yt_sources.py.
Boundary markers: Explicit instructions are provided to the agent in SKILL.md to treat YouTube data as informational data only.
Capability inventory: The agent has access to Bash and Write tools, and executes several local Python scripts.
Sanitization: Content is partially cleaned (e.g., removing pipe characters) before being presented in the candidate table.
[COMMAND_EXECUTION]: The skill uses subprocess.run to call yt-dlp and other internal utility scripts. It also provides a shell command template in SKILL.md that dynamically includes the user's request. There is a potential risk of command injection if the agent does not properly escape the user's input before running the resulting shell command.
[PRIVILEGE_ESCALATION]: The setup instructions in references/setup.md suggest that macOS users grant 'Full Disk Access' to their terminal application. This permission is necessary for the 'Live Mode' feature to read browser cookie databases (e.g., Firefox), but it provides the terminal and the agent with extensive system permissions.
[EXTERNAL_DOWNLOADS]: The skill makes network requests to well-known YouTube services, including RSS feeds and the Google YouTube Data API, to retrieve video recommendations and metadata.

video-lens-recs