The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads AI model weights and checkpoints from well-known public repositories on GitHub (OpenTalker, TencentARC, and xinntao) to facilitate the optional SadTalker avatar functionality. These downloads are performed via a PowerShell script using curl to specific release URLs.
[COMMAND_EXECUTION]: Multiple Python scripts (build_video.py, add_avatar.py, ellipse_overlay.py) utilize subprocess.run to orchestrate the video generation process. This includes calling ffmpeg for media synthesis and invoking local Python scripts for avatar inference within a dedicated virtual environment.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content, such as blog posts fetched from URLs, meeting notes, and pull request diffs. This data is passed to the LLM to generate a structured video script. While this constitutes an attack surface for indirect prompt injection, the skill employs a rigid JSON schema for the output and focuses on summarizing the provided context for video production, which is a standard use case for text-processing agents.
[CREDENTIALS_SAFE]: The YouTube upload integration instructions correctly advise the user to manage authentication tokens and client secrets within a secrets/ directory and include it in .gitignore to prevent accidental exposure.

context-to-video