The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes ffmpeg and ffprobe via the subprocess module to extract video metadata, detect scene cuts, and identify static frames. These invocations use shlex.quote to sanitize file paths and prevent command injection.
[DYNAMIC_EXECUTION]: The orchestrator (scripts/07_orchestrator.py) and the render script (scripts/06_render_diff.py) use importlib.import_module to load pipeline stages dynamically from the local scripts/ directory.
[DYNAMIC_EXECUTION]: To measure fidelity, scripts/06_render_diff.py generates a temporary Python script that uses Playwright to record the generated animation. This script is then executed as a separate process.
[EXTERNAL_DOWNLOADS]: The build_preview_html function in the orchestrator generates an HTML file that references the GSAP animation library from the well-known CDN unpkg.com for preview purposes.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided video content through a Vision Language Model (VLM), which represents a potential injection surface.
Ingestion points: Visual frames extracted from the user's video are provided to the VLM in scripts/04_vlm_analyze.py.
Boundary markers: The system prompt explicitly instructs the model to adhere to visual analysis and ignore timing, while the pipeline enforces a strict JSON schema for all VLM outputs.
Capability inventory: The skill possesses capabilities for file system writes, image/video processing, and subprocess execution.
Sanitization: The skill employs jsonschema to validate the model's response before it is used to generate code, and uses shlex for shell command safety.

reverse-animate