The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides instructions for loading models and datasets from the Hugging Face Hub, which serves as a potential vector for malicious instructions embedded in external artifacts.
Ingestion points: Model loading via from_pretrained() in references/models.md and dataset loading via load_dataset() in references/training.md.
Boundary markers: The skill includes explicit warnings in SKILL.md to review custom code associated with models before setting trust_remote_code=True.
Capability inventory: The skill is configured with Bash, Read, Write, and Edit tools, allowing for file system and command-line operations.
Sanitization: The documentation relies on manual user review of model cards and code as the primary safety mechanism.
[SAFE]: The skill promotes secure credential management by instructing users to avoid hardcoding authentication tokens like HF_TOKEN, recommending instead the use of environment variables, secret managers, or official CLI login tools.
[SAFE]: All documented package dependencies are well-known, industry-standard libraries for machine learning and data science, with specific version pinning to ensure environment stability.

transformers