The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents patterns where untrusted user data (e.g., prompts) is ingested and processed by machine learning models, creating a surface for indirect prompt injection.
Ingestion points: Gradio input components passing data to functions decorated with @spaces.GPU, such as the generate(prompt: str) example in SKILL.md.
Boundary markers: The provided implementation examples do not utilize explicit delimiters or instructions to the model to ignore embedded control sequences.
Capability inventory: The skill uses @spaces.GPU for remote compute and mentions creating temporary files via tempfile.
Sanitization: No explicit sanitization or filtering of the prompt input is demonstrated in the provided code snippets.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The documentation describes the process of installing CUDA-dependent packages from external sources.
Evidence: The skill recommends fetching pre-built binary wheels from the official flash-attention releases on GitHub (e.g., https://github.com/Dao-AILab/flash-attention/releases) and references the kernels-community repository on the Hugging Face Hub to manage runtime compatibility.

huggingface-zerogpu