The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines function templates that ingest untrusted user data for processing by machine learning models, creating a standard surface for potential prompt injection.
Ingestion points: The generate and generate_image function examples in SKILL.md accept a prompt string argument from the user.
Boundary markers: The provided code snippets do not implement delimiters or instructions to the model to ignore potential commands embedded within the input prompt.
Capability inventory: The functions utilize the transformers library and @spaces.GPU for CUDA-accelerated inference tasks.
Sanitization: There are no explicit validation or sanitization steps shown for the input strings before they are passed to the model pipeline.
[EXTERNAL_DOWNLOADS]: The skill provides instructions for managing dependencies and wheels from external sources.
The documentation in references/cuda-and-deps.md suggests downloading pre-built wheels from GitHub releases (e.g., Dao-AILab/flash-attention) and mentions using kernels-community resources on Hugging Face.
It includes a command to query metadata from the official PyPI JSON API using curl piped to a python3 parser for version verification.
These references target well-known development services and official package repositories.

huggingface-zerogpu