The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run in prep_video.py and shell scripts like enumerate_tools.sh to invoke media processing tools such as ffmpeg, whisper, and swift. These tools are necessary for frame extraction, transcription, and OCR processing.\n- [INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface for indirect prompt injection by processing untrusted data sources including video files, audio transcripts, and OCR text. This data is fed into LLM agents for analysis. Evidence chain: (1) Ingestion points: Video files, URLs, Whisper transcripts, and OCR text processed in prep_video.py and ocr.swift. (2) Boundary markers: Prompts in lapian.star and video_understand.star use string interpolation and labels (e.g., {aligned}), but do not provide robust isolation from malicious content. (3) Capability inventory: The skill can execute shell commands via subprocess, perform network downloads, and write to the file system. (4) Sanitization: No sanitization or escaping is performed on the extracted text before it is passed to the AI agents.\n- [EXTERNAL_DOWNLOADS]: The skill facilitates downloading videos from external platforms like Rednote (小红书) and references the author's GitHub repository for the harness workflow tool. These external interactions are essential for the skill's functionality but involve interaction with external infrastructure and untrusted content.\n- [DYNAMIC_EXECUTION]: The skill executes ocr.swift at runtime using the swift command interpreter to perform on-demand OCR processing of extracted video frames.

video-understand