The Agent Skills Directory

[COMMAND_EXECUTION]: Multiple scripts such as render.py, build_timeline.py, and transcribe.py execute ffmpeg and ffprobe via subprocess.run. These operations are standard for media processing and are used to analyze and edit video files as per the skill's core functionality. The commands are constructed using argument lists, which is a safe practice to prevent shell injection.\n- [EXTERNAL_DOWNLOADS]: The skill optionally downloads a sample video from the author's GitHub repository (nextwork-projects/ai-video-editor-cooked). This is a vendor-owned resource used for demonstration and is considered safe.\n- [INDIRECT_PROMPT_INJECTION]: The skill has an ingestion surface for indirect prompt injection because it processes user-provided videos and scripts that are converted to text. \n
Ingestion points: Video files and external scripts provided by the user are ingested and transcribed in SKILL.md and transcribe.py.\n
Boundary markers: Decision-making is based on structured JSON (spans.json) and specific text quotes that are validated against the transcript in build_timeline.py.\n
Capability inventory: The skill can modify the local file system (writing rendered video and metadata) and call external APIs (ElevenLabs for transcription).\n
Sanitization: Text normalization is handled in textnorm.py. The workflow requires explicit user approval of a transcript preview (cut-check.html) before final rendering occurs, providing a strong defense against automated exploitation through malicious content embedded in transcripts.

cut