The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/search.py invokes the yt-dlp command-line utility using subprocess.run. The command is constructed as a list, which prevents shell command injection from user input, but it establishes a dependency on an external tool.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted metadata from YouTube search results.
Ingestion points: Video titles, channel names, and descriptions are fetched from YouTube via the yt-dlp subprocess in scripts/search.py.
Boundary markers: Individual video results are separated in the output by visual Unicode dividers (─) to help the agent distinguish between entries.
Capability inventory: The script's capabilities are limited to executing yt-dlp and printing results; it does not perform file writes, modifications to the system environment, or direct arbitrary network requests.
Sanitization: While the script parses structured JSON output, it does not sanitize or escape the content of video titles or descriptions before displaying them, which may allow malicious instructions embedded in video metadata to influence the agent's downstream behavior.

yt-search