The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides templates and patterns for building Gradio applications that ingest user-provided strings (prompts) for processing by machine learning models.
Ingestion points: Functions like generate(prompt: str) in SKILL.md serve as input points for untrusted data.
Capability inventory: The skill utilizes GPU-based inference capabilities via the transformers and torch libraries.
Boundary markers: The provided code snippets do not demonstrate the use of boundary delimiters or specific "ignore" instructions for the model when processing input.
Sanitization: No explicit input sanitization or validation mechanisms are shown in the basic reference patterns.
[EXTERNAL_DOWNLOADS]: The skill provides guidance on acquiring necessary binary dependencies from external sources to ensure compatibility with ZeroGPU hardware.
Evidence: references/cuda-and-deps.md references downloading pre-built wheels from the official Dao-AILab/flash-attention GitHub repository and using the kernels-community repository on Hugging Face.
Context: These references target well-known research repositories and official community hubs for the purpose of resolving hardware-specific installation constraints.

huggingface-zerogpu