The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches pre-trained machine learning models and weight files from official GitHub repositories of recognized organizations to enable avatar and face-repair functionality.
Evidence: The script scripts/download_models.ps1 downloads model checkpoints from repositories under the OpenTalker, xinntao, and TencentARC GitHub organizations.
[COMMAND_EXECUTION]: Executes FFmpeg for media processing and local Python scripts for AI inference tasks.
Evidence: scripts/build_video.py and scripts/add_avatar.py use subprocess.run to call ffmpeg and local scripts like inference.py using list-based arguments for parameter passing.
[PROMPT_INJECTION]: The skill processes external data (URLs, PR diffs, and documents) to generate video scripts, which presents a surface for indirect prompt injection.
Ingestion points: Untrusted content enters the pipeline via the web_fetch tool (for URLs) or by reading local files such as meeting notes and release logs.
Boundary markers: No explicit delimiters or "ignore instructions" warnings are implemented in the prompts that synthesize input data into the script.json schema.
Capability inventory: The skill can execute shell commands (via FFmpeg), perform network operations (TTS and YouTube APIs), and read/write files within the project workspace.
Sanitization: Data is processed through an LLM summarization step which structures the content into a fixed JSON schema before it is consumed by the rendering scripts.

context-to-video