The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from external sources, including video metadata (titles, descriptions) and transcripts (from native captions or Whisper API output).
Ingestion points: Video metadata is ingested in scripts/download.py, VTT subtitle files in scripts/transcribe.py, and Whisper API responses in scripts/whisper.py.
Boundary markers: The SKILL.md instructs the agent to populate specific  markers within the generated report.md.
Capability inventory: The skill possesses extensive local execution capabilities through Python's subprocess module, which is used to call ffmpeg, ffprobe, and yt-dlp across multiple scripts.
Sanitization: The scripts utilize list-based arguments for subprocess.run to prevent shell injection. Additionally, scripts/download.py implements hardening measures such as resolving local paths to absolute paths and using the -- separator to prevent option injection attacks.
[COMMAND_EXECUTION]: The skill frequently executes local system binaries using Python's subprocess module to perform video and audio processing.
Evidence includes calls to ffprobe and ffmpeg in scripts/frames.py and scripts/hook.py, and yt-dlp in scripts/download.py.
scripts/setup.py also facilitates the installation of these dependencies via Homebrew on macOS or by providing manual installation commands for other platforms.
[EXTERNAL_DOWNLOADS]: The skill uses yt-dlp to download video files and metadata from user-provided URLs.
[DATA_EXFILTRATION]: When native captions are unavailable, the skill extracts audio segments and sends them to Groq or OpenAI Whisper APIs for transcription.
This behavior is clearly documented as a core feature of the skill. The endpoints used (api.groq.com and api.openai.com) are for well-known and trusted services.

watch