The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a driver script (scripts/whisper.py) that manages the execution of local media processing tools such as ffmpeg and whisper-cpp.
The script utilizes subprocess.run with list-based arguments rather than shell strings, effectively preventing shell injection vulnerabilities.
It implements specialized escaping for ffmpeg filtergraphs in the srt-burn subcommand to safely handle input filenames that contain special characters like colons or quotes.
[EXTERNAL_DOWNLOADS]: The skill references standard installation procedures and well-known, trusted repositories for its operational requirements.
Dependencies are installed from official package registries (Homebrew for whisper-cpp and PyPI for faster-whisper).
Model download instructions point to official GGML model repositories on Hugging Face, which is a recognized and reliable source for machine learning models.
[PROMPT_INJECTION]: The skill presents a surface for indirect prompt injection common to all transcription tools, as it processes untrusted audio or video content provided by the user.
Ingestion points: Media files are processed through ffmpeg and Whisper models via the scripts/whisper.py driver.
Boundary markers: The skill does not currently implement delimiters or warnings to isolate transcribed text in the agent's context.
Capability inventory: The skill possesses the ability to execute local binaries for media manipulation and file system operations within the workspace.
Sanitization: Transcribed text is output directly from the model without additional filtering or sanitization.

media-whisper