The Agent Skills Directory

Command Execution: The skill guides the execution of repository management operations via CLI tools and package executors, specifically using npx with @huggingface/hub and the hf command-line utility. These commands are designated for creating repositories and uploading data folders directly to the platform.
Data Exposure Considerations: The documentation includes instructions for archiving local agent session traces located in paths like ~/.claude/projects, ~/.codex/sessions, and ~/.pi/agent/sessions. The skill notes that these files can contain sensitive information such as configuration details, file paths, or credentials, and correctly guides the user to use the --private flag to restrict access to the uploaded datasets.
Indirect Prompt Injection Surface: The skill implements functionality to retrieve and display text data from external, third-party repositories via endpoints like /rows and /search. Because this content is untrusted and externally generated, it presents a potential pipeline for indirect prompt injection if parsed directly into an agent's context.
Ingestion points: Data fields are loaded into the environment from remote dataset shards through the /first-rows, /rows, /search, and /filter API routes.
Boundary markers: No specific text isolation techniques or encapsulation boundaries are explicitly defined in the provided instruction set.
Capability inventory: The system retains the capability to invoke shell actions via curl, npx, and native CLI commands to interact with data layers.
Sanitization: No data validation or structural filtering is detailed for the incoming dataset row attributes prior to presentation.

huggingface-datasets