The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill makes network requests to http://localhost:31337/notify for local voice notifications and https://api.cleanvoice.ai/v2 for optional cloud-based audio processing. These are legitimate use cases for the skill's functionality. The Cleanvoice integration is well-documented and requires a user-provided API key.
[COMMAND_EXECUTION]: Extensive use of shell commands via bun $ and spawnSync to invoke media tools like ffmpeg, ffprobe, whisper, and insanely-fast-whisper. These operations are central to the skill's purpose of audio processing and are performed on user-specified files.
[CREDENTIALS_SAFE]: The skill follows secure practices by instructing users to store sensitive API keys (ANTHROPIC_API_KEY, CLEANVOICE_API_KEY) in a .env file within the ~/.claude/ directory rather than hardcoding them. Tools/Polish.ts contains logic to securely load these from the environment.
[DATA_EXPOSURE]: The skill processes audio files and transcripts. The Polish workflow uploads audio to Cleanvoice.ai. This is a primary feature of the skill, and the documentation explicitly warns users about cloud processing for sensitive content.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted transcription data in Tools/Analyze.ts and interpolates it into a system prompt for Claude to classify edits. While this presents an injection surface, the skill implements a 750-word windowing strategy and uses specific boundary markers (timestamps) which help isolate segment analysis. The resulting classifications only influence ffmpeg cut parameters, limiting the impact of any potential injection.

AudioEditor