The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/transcribe_large.sh is vulnerable to command injection. It uses a bash heredoc to generate and execute a Python script, interpolating the $AUDIO_FILE shell variable directly into a Python string literal without sanitization. Because the heredoc delimiter EOF is unquoted, the shell expands the variable before the Python interpreter sees it. A maliciously crafted filename containing characters like \") followed by Python code could escape the string literal and execute arbitrary Python commands in the context of the user running the script.
[EXTERNAL_DOWNLOADS]: The skill downloads approximately 1.5GB of Whisper model weights from external sources (likely Hugging Face/MLX repositories) during its first execution. While this is necessary for the skill's primary function of local transcription, it involves runtime acquisition of large binary assets.
[COMMAND_EXECUTION]: The skill's documentation and installation instructions suggest installing a macOS LaunchAgent (com.local-whisper.plist) into ~/Library/LaunchAgents/ and loading it with launchctl. This modifies system startup behavior to ensure the transcription daemon runs automatically upon user login, which serves as a persistence mechanism.

whisper-mlx-local