The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run and os.system extensively throughout its reference files (e.g., 3d-cuda-extensions.md, 3d-models.md) to install packages, run compilation scripts, and interact with the hf CLI. These operations are intended for environment setup and deployment of machine learning applications.
[EXTERNAL_DOWNLOADS]: The skill references downloads from various external sources, including specific GitHub repositories for model-specific wheels (e.g., LDYang694/Storages, adithyaxx/flash-attention) and standard Git repositories for model code (e.g., tatsy/torchmcubes, facebookresearch/demucs). These downloads are documented as required dependencies for reproducing official Hugging Face Space environments.
[DYNAMIC_EXECUTION]: The skill describes patterns for runtime compilation of CUDA and C++ extensions using nvcc and g++ (e.g., in 3d-cuda-extensions.md and zerogpu.md). This is a common requirement for high-performance ML models that need custom kernels not available in standard prebuilt packages.
[CREDENTIALS_UNSAFE]: The skill provides instructions on managing sensitive information like Hugging Face API tokens (HF_TOKEN) and SSH keys (e.g., ~/.ssh/hf_dev) to facilitate authenticated access to the Hugging Face Hub and Space development features. These are standard platform procedures.
[INDIRECT_PROMPT_INJECTION]: The skill describes building Gradio applications that ingest user prompts for machine learning tasks (e.g., image or 3D generation). This introduces a potential surface for indirect prompt injection, although the skill includes instructions for managing request flow and process isolation.

huggingface-spaces