The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external audio/video files to generate transcripts and markdown documentation. Since it does not use boundary markers or provide explicit instructions to ignore commands within the transcribed text, it presents a surface for indirect prompt injection attacks. Malicious instructions spoken in a video could be transcribed and subsequently executed by an agent reading the generated files.
Ingestion points: External video and audio files processed via the OpenAI Whisper API in transcribe_video.py.
Boundary markers: Absent. Transcribed text is directly interpolated into Markdown templates such as templates/chapter.md.
Capability inventory: The skill is configured with Bash, Write, and Read capabilities, which increase the potential impact if an agent follows instructions embedded in a transcript.
Sanitization: No sanitization or filtering is performed on the transcript content before it is written to the file system.
[COMMAND_EXECUTION]: The scripts transcribe_video.py, generate_highlights.py, and export_highlight_script.py utilize subprocess.run to call ffmpeg and ffprobe for media processing. These implementations use the list-based argument format, which is a recommended security practice to prevent shell injection. Additionally, generate_highlights.py includes a dedicated escaping function for text overlays to prevent breakage of the FFmpeg filter chain.

video-add-chapters