The Agent Skills Directory

[COMMAND_EXECUTION]: The installation and execution process requires running swift build and swift run to compile and launch a native application from the AiwithDhruv_Voice/WhisperAiwithDhruv directory. This allows for the execution of code contained within the skill's source files.
[EXTERNAL_DOWNLOADS]: The application initiates a network request during its first launch to download machine learning models (approximately 140MB for the base model) required for offline transcription.
[COMMAND_EXECUTION]: The application requires the user to grant 'Accessibility' permissions in macOS System Settings. This high-privilege permission allows the app to monitor and control other applications, which is used to implement the 'auto-type' feature by injecting keystrokes into active text fields.
[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection because it ingests untrusted live audio data that is transcribed into text and provided as an output.
Ingestion points: Live audio captured via the microphone and processed by WhisperKit.
Boundary markers: None documented for the transcribed output string.
Capability inventory: The skill uses swift build and swift run for execution, and Accessibility permissions for system-wide text injection.
Sanitization: The skill documentation mentions filtering for Whisper-specific artifacts (hallucinations), but it does not appear to sanitize the transcribed text for potential prompt injection or malicious instructions that could affect downstream agent tasks.

whisper-voice