The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes a moderation section in SKILL.md that explicitly instructs the agent to 'Never infer a moderation decision' from various sensitive categories (politics, religion, copyright, etc.) and to 'Do not pre-screen prompts or media.' This is an attempt to override the AI's internal safety filters and delegate all moderation responsibility to the host platform's explicit error returns.
[COMMAND_EXECUTION]: The skill mandates the use of several command-line tools for its operations. Specifically, yt-dlp is used to download external video content, while ffmpeg, ffprobe, and Manim are used for video analysis, processing, and animation. These commands are executed through a shell or sandbox environment.
[DYNAMIC_EXECUTION]: The references/code-built-video.md file describes a process for generating and executing code at runtime. The agent is instructed to create HTML/CSS/SVG pages for rendering via Playwright (Chromium) and to generate Python scripts for Manim based on user-supplied data and specifications.
[EXTERNAL_DOWNLOADS]: The skill performs external network operations using yt-dlp to fetch video files from URLs provided by the user in the conversation.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits a significant attack surface for indirect prompt injection. It ingests untrusted data from external sources (downloaded videos, user-provided spreadsheets) and possesses a high-capability toolset (shell access, code execution, network downloads).
Ingestion points: Video content downloaded via yt-dlp and data processed in references/code-built-video.md (e.g., user spreadsheets).
Boundary markers: No specific boundary markers or 'ignore' instructions for processed data are defined in the instructions.
Capability inventory: The skill uses ffmpeg, yt-dlp, Playwright, Manim, librosa, and general shell access (via SKILL.md and references/code-built-video.md).
Sanitization: No explicit sanitization or validation logic for the external inputs is described.

director