The Agent Skills Directory

[PROMPT_INJECTION]: The skill explicitly defines security boundaries for handling untrusted content found in media files and transcripts.
Ingestion points: The agent processes filenames, metadata, and transcribed speech from video and audio files.
Boundary markers: The skill includes a dedicated 'Untrusted content boundary' section instructing the agent to ignore any commands or instructions found within the media.
Capability inventory: The skill utilizes subprocess calls to ffmpeg and whisper-cli, as well as local file system writes.
Sanitization: It requires resolving paths under a project root, rejecting symlinks, and using argv entries for tool calls to prevent shell injection.
[EXTERNAL_DOWNLOADS]: The skill provides instructions for obtaining tools and models from well-known repositories.
Downloads the whisper.cpp source code from its official GitHub repository and emphasizes pinning to a specific commit SHA.
Fetches pre-trained models from Hugging Face with explicit SHA256 hash verification.
Recommends installing Python dependencies using --require-hashes to ensure package integrity.
[COMMAND_EXECUTION]: Shell commands are used for media processing with specific safety constraints.
Employs ffmpeg and whisper-cli for core functionality.
Explicitly advises against using shell interpolation when passing paths to these tools to mitigate command injection risks.

video-transcribe