The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted data from external video sources, such as transcripts, subtitles, and metadata, which creates a surface for indirect prompt injection.
Ingestion points: Data enters the agent's context through scripts/watch.py, which processes external video content and transcription results.
Boundary markers: No explicit delimiters or instructions are used to separate untrusted transcript text from agent instructions.
Capability inventory: The agent has access to Bash and Read tools, providing a capability surface if successfully injected.
Sanitization: Subtitle text is extracted but not sanitized or filtered for potential malicious instructions.
[COMMAND_EXECUTION]: The skill executes external binaries to process media files.
Evidence: scripts/download.py, scripts/frames.py, and scripts/setup.py utilize subprocess.run() to invoke ffmpeg, ffprobe, and yt-dlp.
Arguments are passed as lists, which follows security best practices to prevent shell injection.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of external tools and media content.
Evidence: scripts/setup.py manages the installation of ffmpeg and yt-dlp via Homebrew (macOS) or WinGet (Windows).
Evidence: scripts/download.py uses yt-dlp to fetch video and subtitle files from user-provided URLs.
[DATA_EXFILTRATION]: The skill sends audio data extracted from videos to third-party transcription services.
Evidence: scripts/whisper_api.py uploads audio files to api.groq.com (Groq) and api.openai.com (OpenAI).
This behavior is essential for the skill's functionality and utilizes well-known service endpoints.

watch