The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill uses uvx to download and run the mlx-whisper and openai-whisper packages from PyPI. It also downloads transcription models (approximately 1.5 GB) from public repositories during the first execution.
[COMMAND_EXECUTION]: The skill executes shell commands via Bash to perform audio conversion using ffmpeg and transcription using uvx. It uses mktemp to create temporary directories for output.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted audio and video files provided by the user or downloaded from external sources. The resulting transcript is delivered back to the agent's context. While the transcript could contain malicious instructions (spoken content designed to manipulate the AI), the skill itself does not have elevated privileges that would make this uniquely dangerous beyond standard LLM interaction risks.
Ingestion points: Audio/video files processed by mlx-whisper or openai-whisper (SKILL.md).
Boundary markers: None explicitly defined for the resulting transcript text.
Capability inventory: File read/write, shell command execution via Bash.
Sanitization: None specified for the text generated from audio.

transcribe