The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes the external tool ffmpeg via subprocess.run to normalize input audio files to a 16kHz mono WAV format. The execution passes arguments as a safe list structure without invoking a shell environment, reducing the possibility of command injection.
[INDIRECT_PROMPT_INJECTION]: The skill processes external, potentially untrusted audio files and transcribes their spoken content into plain text, introducing an inherent indirect prompt injection risk surface if the transcribed text contains adversarial instructions directed at the downstream agent.
Ingestion points: The script scripts/stt.py ingests user-supplied audio tracks through the audio_file command-line argument.
Boundary markers: There are no explicit boundaries, delimiters, or safety instructions wrapping the transcribed text output.
Capability inventory: The transcription output is written to standard output for consumption by the agent environment, and can optionally trigger an outbound network call via requests.put to a user-configured Matrix homeserver.
Sanitization: The script performs no content validation or text sanitization on the resulting transcription string before passing it downstream.

ez-stt