The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external media files and transcription data (SRT files) to determine edit points. This creates a surface for indirect prompt injection, where malicious instructions embedded in the audio or resulting transcript could attempt to manipulate the agent's behavior.
Ingestion points: media files (mp4, wav), transcript files (SRT).
Boundary markers: The skill does not explicitly describe the use of delimiters or warnings to ignore instructions embedded within transcripts.
Capability inventory: The agent has access to vcut commands, EDL building, and rendering tools, which involve significant file system operations.
Sanitization: No explicit sanitization of transcript content for malicious command strings is mentioned.
[EXTERNAL_DOWNLOADS]: The skill documentation and accompanying Python script describe a setup process (vcut setup classifier) that downloads a 320MB machine learning model into the ~/.vcut/panns directory. While this is part of the core functionality, it represents an external binary dependency.
[COMMAND_EXECUTION]: The skill uses subprocess.run in the scripts/non-speech.py utility to call ffmpeg for audio processing. The command is constructed using a list of arguments, which mitigates simple shell injection risks. The skill also utilizes various vcut subcommands to manage video editing sessions and renders.

core