The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/projectlib.py dynamically loads and executes a Python module (motion_graphics_plan.py) from a sibling skill directory using importlib.util with a path computed relative to the library's location. This allows runtime execution of code from a file whose path is determined at execution time.
[INDIRECT_PROMPT_INJECTION]: The skill's primary workflow involves transcribing video audio and providing the resulting text to the AI agent for summarization. This creates an attack surface where malicious instructions embedded in the video's audio could be interpreted by the AI as commands, potentially bypassing intended constraints.
Ingestion points: Transcription text generated in work/understand/transcript.json is read by the agent for summarization.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are used when the agent processes the transcript.
Capability inventory: The agent has the ability to execute shell commands (via ffmpeg/ffprobe) and perform file system operations.
Sanitization: No sanitization or filtering of the transcript content is performed before it is presented to the agent.
[COMMAND_EXECUTION]: The skill extensively executes external binaries such as ffmpeg and ffprobe using the subprocess module. While these calls use argument lists rather than a shell, the arguments are constructed from data stored in various JSON configuration files (project.json, media.json, etc.), which could lead to argument injection if those files are tampered with.
[EXTERNAL_DOWNLOADS]: The transcription process in scripts/transcribe.py utilizes the faster-whisper library, which fetches AI models from Hugging Face's official repositories at runtime. This is standard behavior for the library to retrieve the required model weights.

video-understand