The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to execute various external binaries and system tools. This includes ffmpeg and ffprobe for video processing, magick and identify (ImageMagick) for image manipulation, and package managers for environment setup.
Evidence: Found in scripts/extract_reference_frames.py (FFmpeg), scripts/remove_chroma_key.py (ImageMagick), and scripts/setup_asset_pipeline.py (Pip and system package managers).
[EXTERNAL_DOWNLOADS]: The skill contains automation for downloading and installing external software, specifically the Pillow Python library and system-level media tools.
Evidence: scripts/setup_asset_pipeline.py manages the installation of dependencies using uv, pip, or system-level managers like brew, apt-get, or winget.
Context: The instructions in references/asset-acquisition.md explicitly gate these operations behind user authorization requirements.
[PRIVILEGE_ESCALATION]: The setup script includes commands that utilize sudo to install system-wide dependencies on Linux platforms.
Evidence: scripts/setup_asset_pipeline.py contains the imagemagick_install_command function which uses sudo for apt-get, dnf, and pacman operations.
Context: This is a legitimate requirement for system-level tool installation, and the skill instructions mandate explicit user approval before performing administrator-level actions.
[DYNAMIC_EXECUTION]: The scripts/remove_chroma_key.py script includes logic to dynamically re-execute itself using a different Python interpreter path identified at runtime (specifically to transition into an isolated virtual environment).
Evidence: The reexec_isolated function in scripts/remove_chroma_key.py uses subprocess.run to restart the script using a path determined by venv_python().
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection as it is designed to ingest and process untrusted external media (video and images) and integrate content from external AI services (Magnific, Higgsfield).
Ingestion points: scripts/extract_reference_frames.py (user-provided MP4 files), references/asset-acquisition.md (external image assets).
Boundary markers: The skill employs a multi-stage validation workflow (Direction, Composition, Rhythm, Craft), though it lacks specific technical sanitization for media metadata.
Capability inventory: The skill possesses file-write access, command execution (FFmpeg/ImageMagick), and network interaction capabilities via MCPs.
Sanitization: Not explicitly implemented for media stream metadata.
Severity: LOW.

balise-motion-graphic