The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes several Python scripts in the scripts/ directory that use the subprocess module to perform media analysis:
probe_video.py executes ffprobe to retrieve video metadata such as duration and stream information.
sample_frames.py uses ffmpeg and ffprobe to extract specific frames for visual inspection.
check_repo.py runs unit tests using the Python unittest module.
These executions are benign, necessary for the skill's functionality, and use safe practices like absolute path resolution via shutil.which and argument arrays instead of shell strings.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (video files and user narratives), which represents a potential attack surface. However, it implements strong mitigation through a mandatory 'Narrative Confirmation' step:
Ingestion points: Untrusted media files and user-supplied descriptions in SKILL.md.
Boundary markers: The agent is explicitly instructed to summarize its findings and wait for user approval (interaction status WAITING_FOR_CONFIRMATION) before generating executable prompts.
Capability inventory: The skill uses Pillow for image generation and ffmpeg/ffprobe for media processing.
Sanitization: By requiring human alignment on the narrative and intent, the skill prevents the agent from autonomously acting on malicious instructions that might be embedded in video metadata or content.
[EXTERNAL_DOWNLOADS]: The README.md and sponsorship documentation reference an external service, whatstoken.ai, via affiliate links. These are identified as sponsorship for an AI API platform and do not involve the automated download or execution of remote code or scripts.

minimax-h3-video-reverse