The Agent Skills Directory

[COMMAND_EXECUTION] (MEDIUM): Multiple scripts (transcribe_video.py, generate_highlights.py, export_highlight_script.py) execute external binaries ffmpeg and ffprobe using subprocess.run. Although arguments are passed as a list to prevent shell injection, generate_highlights.py dynamically constructs a complex filter_complex string using text extracted from transcript markdown files. An attacker providing a maliciously crafted transcript could attempt to exploit the FFmpeg drawtext filter, although the skill attempts basic escaping of characters like colons and quotes.
[EXTERNAL_DOWNLOADS] (LOW): The skill requires system-level installation of ffmpeg and ffprobe. While these are standard tools, their presence is a prerequisite for the skill's execution environment.
[INDIRECT_PROMPT_INJECTION] (LOW): The skill processes untrusted audio content (via Whisper) and user-edited markdown transcripts. While the current scripts treat this data as strings for regex and FFmpeg overlays without allowing it to influence control flow, a downstream agent reading the 'cleaned' transcripts could be susceptible to instructions embedded in the video audio.
[METADATA_POISONING] (LOW): SKILL.md references generate_docs.py as a core step in the workflow, but this file is missing from the package. This suggests either an incomplete skill distribution or misleading documentation.

video-add-chapters