The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the hf CLI for repository management (hf repos create), authentication (hf auth login), and deployment (hf upload). It also utilizes subprocess.run and os.system within Python scripts to handle package installation and runtime compilation of C++ extensions, which are standard practices for building complex ML applications on Hugging Face ZeroGPU.
[EXTERNAL_DOWNLOADS]: Dependencies are fetched from trusted and well-known sources including PyPI, the official Hugging Face Hub (models and datasets), and GitHub. The skill provides recipes for installing prebuilt wheels from the multimodalart/zerogpu-blackwell-wheels dataset on Hugging Face to optimize build times and reliability.
[CREDENTIALS_SAFE]: Instructions correctly guide the user to handle sensitive information using Hugging Face Secrets (--secrets) or environment variables inside the Space, rather than hardcoding credentials. Authentication is performed via standard OAuth flows provided by the hf CLI.
[DYNAMIC_EXECUTION]: The skill documents legitimate runtime execution patterns such as JIT compilation for CUDA extensions using nvcc and g++, as well as Ahead-of-Time Inductor (AoTI) compilation for optimizing model inference. These are advanced development techniques required by the specific hardware constraints of the platform.
[INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface by building Gradio applications that ingest untrusted user data (text, images) which is then processed by machine learning models. However, this is inherent to the purpose of the skill and the risk is mitigated by platform-level guardrails.

huggingface-spaces