The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes system binaries ffmpeg and ffprobe to perform video rendering, frame extraction, and metadata probing. This is a core part of its subtitle-burning functionality.
Evidence: scripts/render.sh executes ffmpeg for video encoding; scripts/gen_ass.py and scripts/match_style_from_image.py use subprocess.check_output() to invoke ffprobe for video resolution detection.
[INDIRECT_PROMPT_INJECTION]: The skill processes external media files (video/audio) to extract transcription data, creating a potential surface for indirect injection via audio content.
Ingestion points: scripts/extract_words.py ingests user-provided audio/video files for transcription.
Boundary markers: No explicit instruction delimiters or boundary warnings are present in the transcription workflow.
Capability inventory: The skill can execute shell commands via subprocess and ffmpeg and perform file-write operations for subtitle and video files.
Sanitization: Data is processed through the mlx-whisper model and serialized via the standard json library, reducing the risk of malicious structural injection into downstream scripts.
[EXTERNAL_DOWNLOADS]: The skill downloads pre-trained machine learning models required for transcription.
Evidence: scripts/extract_words.py fetches models from the mlx-community repository on Hugging Face (e.g., whisper-large-v3-turbo). This targets a well-known community organization providing optimized models for Apple Silicon.

karaoke-captions