The Agent Skills Directory

[SAFE]: The skill is entirely composed of documentation and instructional code snippets for PyTorch development. No malicious patterns, obfuscation, or unauthorized access attempts were identified.
[INDIRECT_PROMPT_INJECTION]: The skill defines patterns for data ingestion, which is a standard surface for indirect injection in machine learning contexts.
Ingestion points: The ImageDataset class in SKILL.md defines a mechanism for reading image files from a local directory.
Boundary markers: Not present in the educational code snippets.
Capability inventory: The skill demonstrates file reading (Image.open), model checkpointing (torch.save/torch.load), and JIT compilation (torch.compile).
Sanitization: The skill mitigates risks by explicitly recommending weights_only=True when using torch.load, which prevents arbitrary code execution during model deserialization.

pytorch-patterns