The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes ffmpeg and ffprobe via subprocess.run across several scripts (render.py, timeline_view.py, self_eval.py) to process video and audio. These executions use list-based arguments, which prevents shell injection, and the commands are restricted to the skill's core functionality.
[EXTERNAL_DOWNLOADS]: The transcribe_local.py script utilizes the faster-whisper library, which downloads pre-trained AI models from Hugging Face. The skill also mentions optional support for yt-dlp to download remote video sources. These are well-known services used for their intended purposes.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted audio data and transcribes it into text that the agent interprets.
Ingestion points: Audio content from user-provided video files is transcribed into takes_packed.md, which serves as the primary input for the agent's reasoning.
Boundary markers: The takes_packed.md file uses Markdown headers and timestamped blocks to delimit segments.
Capability inventory: The skill can execute ffmpeg commands, write to the local file system, and potentially download external content via yt-dlp.
Sanitization: The skill does not currently sanitize or escape the transcribed text content to prevent the agent from following instructions embedded within the audio.

video-cut