The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses execFile to run yt-dlp for downloading audio and ffmpeg/ffprobe for media processing. These commands are integral to the skill's primary purpose and handle user-supplied URLs or file paths.
Evidence in index.ts: Calls to execFileAsync('yt-dlp', ...) and execFileAsync('ffmpeg', ...).
[EXTERNAL_DOWNLOADS]: The skill downloads audio content from YouTube, Instagram, and TikTok using the well-known utility yt-dlp. This is the core functionality of the skill.
Evidence in index.ts: downloadAudio function utilizes yt-dlp to fetch media from external platforms.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted transcription text and passes it to an LLM for categorization. While this presents an attack surface where a transcript could contain instructions, the skill implements mitigations and the risk is low.
Ingestion points: External media content is transcribed into text in index.ts.
Boundary markers: The prompt in classifyAndOrganize uses clear section headers like NEW TRANSCRIPT TO ORGANIZE: to separate data from instructions.
Capability inventory: The skill has execFile and fs.writeFile capabilities.
Sanitization: The transcript is truncated and sampled in truncateTranscript to fit token limits, providing some implicit filtering.
[SAFE]: The skill uses official SDKs for Anthropic and ElevenLabs. API keys are managed via a .env file, which is a standard and safe practice for secret management. Access to ~/Documents/transcripts is consistent with the stated purpose of building a knowledge library.

transcribee