The Agent Skills Directory

[DYNAMIC_EXECUTION]: The render_video.py script implements a self-healing render loop. When a Manim scene fails to render, the script sends the error traceback and original source code to an LLM, which returns a patched version of the script. The skill then overwrites the local file with this LLM-generated content and re-executes the Python process via subprocess. This feedback loop presents a risk as it executes code generated by a remote model based on user-influenced context.
[INDIRECT_PROMPT_INJECTION]: The skill takes user-provided concept descriptions and interpolates them into templates for LLM prompts in plan_storyboard.py and _fixup_client.py. These templates lack boundary markers or sanitization logic, creating a surface where untrusted input can influence the instructions sent to the model that generates the executable code.
[COMMAND_EXECUTION]: Subprocess calls are used throughout the skill to interact with system tools. render_video.py executes manim to render scenes, while add_audio.py and critic_pass.py use ffmpeg and ffprobe for video processing and frame extraction. These tools are invoked with arguments derived from the skill's logic and LLM outputs.
[EXTERNAL_DOWNLOADS]: The skill resolves and downloads external image assets via the IconFinder API. The fetch_assets.py script makes network requests to api.iconfinder.com and subsequently downloads files from the URLs returned by the search service using urllib.request.urlopen.

concept-to-video