The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes video transcripts (untrusted data) and interpolates them into a prompt (VIRAL_DETECTION_PROMPT) used for AI analysis without clear boundary markers or sanitization.
Ingestion points: clips_machine.py
transcribe_with_timestamps extracts text from video files or URLs using whisper-cpp.
Boundary markers: Absent. The transcript content is directly embedded into the instructions for the viral moment detection prompt.
Capability inventory: Extensive use of subprocess.run to call ffmpeg, yt-dlp, and whisper-cpp binaries (in clips_machine.py).
Sanitization: Absent. The transcript text is interpolated into the prompt without filtering or escaping designed to prevent instructions hidden in audio from being interpreted as commands by the AI.
[COMMAND_EXECUTION]: The script makes extensive use of subprocess.run to call system binaries to perform video processing tasks.
Evidence: Functions such as get_video_duration, download_video, transcribe_with_timestamps, cut_video, crop_to_vertical, and create_styled_captions all execute external shell commands via argument lists. While argument lists mitigate basic shell injection, the reliance on these tools represents a significant execution surface.
[EXTERNAL_DOWNLOADS]: The skill uses yt-dlp to download video content from external URLs.
Evidence: The download_video function in clips_machine.py fetches data from user-provided URLs. The script implements an ALLOWED_DOMAINS whitelist to restrict downloads to known video hosting platforms, which is a positive security control.

clips-machine