The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided Python code and data to provide optimization recommendations, creating an inherent attack surface for indirect prompt injection.
Ingestion points: User-provided numerical and scientific Python code (NumPy, pandas, SciPy, etc.) and dataset descriptions in the conversation context (SKILL.md, references/code_transformation_patterns.md).
Boundary markers: The instructions do not specify explicit delimiters or warnings for the agent to ignore instructions embedded within user code or data comments.
Capability inventory: The skill includes instructions for package installation via uv add (references/installation.md), shell command execution for profiling (e.g., nsys, ncu), and Python code execution for benchmarking and data processing.
Sanitization: No instructions are present for sanitizing or escaping content from user-provided code snippets before processing.
[EXTERNAL_DOWNLOADS]: The skill provides instructions to install numerous GPU-accelerated libraries from official package registries.
Evidence: Installation commands in references/installation.md use uv add with --extra-index-url=https://pypi.nvidia.com for packages such as cugraph-cu12, cuml-cu12, and cuvs-cu12. These target official NVIDIA infrastructure, which is a well-known and trusted service.
[DYNAMIC_EXECUTION]: The skill utilizes Just-In-Time (JIT) compilation and data serialization techniques common in high-performance computing.
Evidence: Extensive use of JIT compilation via Numba-CUDA (@cuda.jit) and NVIDIA Warp (@wp.kernel) is documented in references/numba.md and references/warp.md. Additionally, references/cuml.md provides patterns for model serialization using the standard pickle module.

optimize-for-gpu