The Agent Skills Directory

[COMMAND_EXECUTION]: The skill contains several shell scripts for managing the OmniVoice Studio backend and recording reference audio. Executed commands include ffmpeg for audio capture and processing, uv and uvicorn for service execution, and kill/lsof for process lifecycle management. On macOS, it also utilizes say and afplay to provide audible cues to the user during the microphone recording process.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of the OmniVoice Studio source code from GitHub and the 2.4 GB diffusion model from HuggingFace. These resources are required for the local backend to function and originate from well-known repositories associated with the tool.
[DATA_EXFILTRATION]: The record-reference.sh script utilizes ffmpeg to capture audio from the system microphone. This data is stored temporarily and then uploaded to the local backend server at 127.0.0.1:3900 for creating voice profiles. The script ensures transparency by using audible announcements to inform the user when recording is active.
[PROMPT_INJECTION]: The generate_speech tool processes arbitrary text through the synthesis engine, creating an indirect prompt injection surface.
Ingestion points: The text parameter of the generate_speech tool in SKILL.md allows the agent to process external content.
Boundary markers: No specific delimiters or instructions are used to separate the input text from the synthesis instructions.
Capability inventory: The skill possesses the capability to record audio, write files, and execute shell commands via its included scripts.
Sanitization: The input text is passed to the local synthesis server without sanitization or filtering of potential embedded instructions.

omnivoice