The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the official hf CLI for managing Hugging Face Spaces and gradio_client for API verification. These are standard platform tools for the intended use case.
[EXTERNAL_DOWNLOADS]: Fetches prebuilt CUDA wheels from Hugging Face datasets and research packages from trusted GitHub organizations. These downloads are appropriate for the specialized machine learning hardware targets mentioned, such as ZeroGPU Blackwell.
[REMOTE_CODE_EXECUTION]: Includes instructions for the agent to install packages directly from source using subprocess.run to resolve environment issues. These operations target trusted repositories and are documented as troubleshooting steps for specific dependency conflicts.
[INDIRECT_PROMPT_INJECTION]: The skill defines an attack surface for indirect prompt injection as it facilitates building Gradio applications that ingest untrusted user data (e.g., text and image inputs).
Ingestion points: Gradio input components in app.py.
Boundary markers: Not explicitly enforced in provided code snippets, relying on standard Gradio behavior.
Capability inventory: Project includes file system access for model loading, subprocess calls for CLI operations, and network access for API client testing.
Sanitization: Sanitization of external content is not explicitly addressed in the minimal examples.
[SAFE]: Authentication and credential management follow best practices by using official CLI methods and platform-native secret environment variables, avoiding hardcoded keys or insecure persistence.

huggingface-spaces