The Agent Skills Directory

[COMMAND_EXECUTION]: The script uses subprocess.run to call the macOS mdls utility for extracting video duration metadata. The command is executed using a list of arguments rather than a shell string, which effectively prevents shell injection vulnerabilities.
[EXTERNAL_DOWNLOADS]: The scripts/run_video_analysis.sh script automates the installation of the google-genai and yt-dlp packages. These are well-known, legitimate libraries fetched from standard registries to support the skill's core functionality.
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as it ingests untrusted data from external sources and interpolates it into LLM prompts.
Ingestion points: Untrusted data enters the context via YouTube metadata (extracted by yt-dlp) and the contents of a user-specified file provided through the --prompt-file argument.
Boundary markers: The skill does not use specific delimiters or instructions to isolate external content from the system prompt.
Capability inventory: The skill has the capability to execute system commands (mdls), perform network operations (Gemini API), and write files to disk (Path.write_text).
Sanitization: There is no evidence of content sanitization or validation for the ingested data before it is sent to the model.

google-video-understanding