The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill implements an automated 'auto-fix' and 'critic' loop in scripts/render_video.py and scripts/critic_pass.py. These scripts receive Python source code generated by a remote LLM (Anthropic API), write that code to the local filesystem, and then execute it via a subprocess call to the Manim engine. This creates a risk where malicious code could be generated and executed if the LLM's instructions are compromised.
[COMMAND_EXECUTION]: Multiple scripts including scripts/render_video.py, scripts/add_audio.py, and scripts/critic_pass.py use subprocess.run to execute external system binaries such as manim, ffmpeg, and ffprobe.
[EXTERNAL_DOWNLOADS]: The scripts/fetch_assets.py script downloads external media assets from the IconFinder API (api.iconfinder.com) based on storyboard requirements and stores them in the local temporary directory.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it processes untrusted user concepts into structured storyboards that subsequently drive the code generation process.
Ingestion points: The user-provided concept text enters the pipeline via scripts/plan_storyboard.py.
Boundary markers: No boundary markers or instruction-following delimiters are present in the references/code2video/planner.md template.
Capability inventory: The skill can perform file system writes and execute arbitrary Python code through the Manim engine as part of its rendering and fixup cycles.
Sanitization: There is no evidence of sanitization or safety filtering applied to the user input before it is interpolated into LLM prompts.

concept-to-video