The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted content from the Hugging Face Hub, such as dataset rows and repository metadata, which could contain malicious instructions designed to influence the agent.\n
Ingestion points: Model and dataset repository cards, file trees, and dataset rows retrieved via the Dataset Viewer API.\n
Boundary markers: Directives like "Inspect before transfer" and the requirement to verify facts against live Hub data serve as boundary markers.\n
Capability inventory: The agent can execute hf CLI commands for file operations, repository management, and compute job control.\n
Sanitization: The skill mandates explicit user authorization for all external mutations, visibility changes, and paid resource consumption.\n- [EXTERNAL_DOWNLOADS]: Retrieves models, datasets, and documentation from the Hugging Face Hub.\n
Evidence: The skill interacts with official domains (huggingface.co) and refers to the huggingface_hub package, which are trusted and well-known services.\n- [COMMAND_EXECUTION]: Uses the local hf CLI to perform repository and job operations.\n
Evidence: Extensive use of commands like hf download, hf upload, and hf jobs list to interact with the Hub.

huggingface