The Agent Skills Directory

[COMMAND_EXECUTION]: Template files designed for user modification (templates/basic_scene.py, templates/3d_scene.py, and templates/math_scene.py) include os.system() calls within if __name__ == "__main__": blocks. These are implemented to automatically invoke the manimgl command-line tool on the file itself, allowing users to render animations by running the script directly with Python.
[EXTERNAL_DOWNLOADS]: The skill references image textures hosted on Wikimedia Commons (e.g., in rules/3d.md and templates/3d_scene.py) which are downloaded at runtime by the TexturedSurface class to skin 3D objects like spheres.
[INDIRECT_PROMPT_INJECTION]: The skill presents a surface for indirect prompt injection due to its combination of untrusted data handling and powerful capabilities.
Ingestion points: The agent is instructed to read, copy, and modify template files (templates/*.py) and rulesets provided in the skill.
Boundary markers: Absent; there are no explicit instructions for the agent to ignore embedded commands within the code it processes.
Capability inventory: The provided scripts include shell command execution via os.system and demonstrate the use of self.embed(), which drops the user into an interactive IPython shell allowing arbitrary code execution.
Sanitization: Absent; the skill does not suggest any filtering or escaping of the content before the agent interpolates it into its workflow.

manimgl-best-practices