The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes several utility scripts (render_video.py, add_audio.py, critic_pass.py) that interface with system tools like manim, ffmpeg, and ffprobe via subprocess.run. These calls are executed using argument lists, which effectively prevents shell injection vulnerabilities.
[EXTERNAL_DOWNLOADS]: The scripts/fetch_assets.py script retrieves visual icons and images from the IconFinder API (api.iconfinder.com). It uses urllib to securely fetch assets based on storyboard requirements.
[DYNAMIC_EXECUTION]: The core functionality relies on an agentic pipeline where an LLM generates Python code (Manim scenes) from a storyboard. An auto-fix loop in scripts/render_video.py and scripts/_fixup_client.py uses traceback data to patch code at runtime if rendering fails. This is a standard and necessary behavior for the skill's primary purpose of programmatic video generation.
[PROMPT_INJECTION]: The skill processes untrusted user concept text which is passed through several LLM and VLM stages. This presents a potential surface for indirect prompt injection; however, the skill implements structured storyboard validation and utilizes prompt templates with clear boundary markers to manage the flow of user-provided data.
Ingestion points: User concept text enters the pipeline via scripts/plan_storyboard.py.
Boundary markers: Prompt templates in references/code2video/ use markdown headers and explicit role instructions to separate system instructions from processed data.
Capability inventory: The generated animation code is executed locally by the Manim engine, which is restricted to animation primitives.
Sanitization: The skill performs comprehensive JSON schema validation on the generated storyboards before they are used to produce executable Python code.

concept-to-video