The Agent Skills Directory

[DYNAMIC_EXECUTION]: The scripts/render_video.py script implements an automated fixup loop. If a Manim render fails, the script sends the source code and error traceback to a model, receives a code patch, applies it to the scene file, and re-executes the render command.
[DYNAMIC_EXECUTION]: The test suite in tests/test_scene_templates.py utilizes exec() and compile() to execute and verify the correctness of generated animation source code against a stub module.
[COMMAND_EXECUTION]: The skill makes extensive use of subprocess.run() to invoke system binaries such as manim for video rendering, and ffmpeg and ffprobe for video processing, duration verification, and audio overlay.
[EXTERNAL_DOWNLOADS]: The scripts/fetch_assets.py utility downloads visual assets from IconFinder's official API for inclusion in generated animations.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface for indirect injection as it processes user-supplied configuration data (JSON/TOML) in scripts/build_scene.py to drive Python code generation. Evidence of this chain includes:
Ingestion points: User-provided config files are loaded in scripts/build_scene.py and passed to the generation logic.
Boundary markers: The Manim templates in references/templates/ use explicit # ── Config and rule markers to delimit the injected configuration block.
Capability inventory: The skill possesses capabilities for subprocess execution (manim, ffmpeg, ffprobe) across multiple scripts and network access for asset fetching.
Sanitization: The scripts/scene_templates.py module attempts to sanitize user input by processing text through repr() when generating Python string literals.

animate