The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions involve compiling and running local Swift code within the project directory.
Evidence: swift build and swift run WhisperAiwithDhruv commands located in the Process section of SKILL.md.
[PRIVILEGE_ESCALATION]: The skill requires manual granting of high-level macOS permissions to the application to enable its core features.
Evidence: Step 3 in SKILL.md instructs the user to grant Microphone and Accessibility permissions in System Settings.
Accessibility permissions allow the software to control and simulate input in other applications, which is a significant privilege level.
[INDIRECT_PROMPT_INJECTION]: The application ingests live microphone audio, creating a data entry point that could be used to deliver malicious instructions to an agent processing the transcripts.
Ingestion points: Microphone audio stream captured in real-time.
Boundary markers: None present; the instructions do not specify the use of delimiters or markers to isolate transcription output.
Capability inventory: The skill outputs transcribed_text and possesses the capability to auto-type text via Accessibility APIs.
Sanitization: The documentation mentions filtering common Whisper hallucinations, but does not identify safeguards against adversarial spoken commands.
[EXTERNAL_DOWNLOADS]: The skill triggers an external download of the Whisper model during its first launch.
Evidence: The 'First launch setup' section in SKILL.md specifies a ~140MB download for the base model.
This download is part of the standard initialization for WhisperKit, which typically fetches assets from well-known model repositories.

whisper-voice