The Agent Skills Directory

[SAFE]: The skill documents a legitimate development process for integrating video diffusion models.
[INDIRECT_PROMPT_INJECTION]: The skill defines a surface for processing external data (upstream repositories, model cards, and checkpoint metadata).
Ingestion points: The agent is instructed to read model cards, upstream repository configurations, and checkpoint index files (.safetensors.index.json) to scope the integration task.
Boundary markers: The skill provides a structured 'map' and 'route' metaphor, directing the agent to specific architectural references to bound the context.
Capability inventory: The integration process involves file system writes (creating config.py, runner.py), shell execution (running uv, pytest, ruff), and network access for downloading model weights.
Sanitization: The instructions explicitly recommend using safetensors and the weights_only=True flag for torch.load(), which are standard security measures to prevent malicious code execution from model files.
[EXTERNAL_DOWNLOADS]: The skill describes downloading model checkpoints from well-known repositories, specifically HuggingFace repo IDs. These operations are part of the intended model integration workflow.
[COMMAND_EXECUTION]: The skill uses standard development tooling for Python environments, including pip, uv, and pytest. These commands are standard for the described development and testing tasks.

integrate-a-model