The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides commands to install official NVIDIA and RAPIDS packages (e.g., cudf, cupy, cuml, cugraph) from PyPI and pypi.nvidia.com. These are established, well-known services for GPU computing.
[DYNAMIC_EXECUTION]: The skill documents the use of Just-In-Time (JIT) compilation technologies such as Numba-CUDA and NVIDIA Warp. These are legitimate tools used to compile Python functions into GPU kernels at runtime to achieve performance gains in scientific computing.
[INDIRECT_PROMPT_INJECTION]: The skill includes functionality for the agent to ingest and process external data files from local storage, S3, or HTTP (via KvikIO and cuDF). While this creates a processing surface for external data, it is the primary intended use case for a data science and optimization skill.
Ingestion points: Data entering through cudf.read_csv, cudf.read_parquet, and kvikio.RemoteFile (found in SKILL.md and references/kvikio.md).
Boundary markers: Not explicitly defined in the instructional code snippets.
Capability inventory: File writing (kvikio.CuFile), network access (kvikio.RemoteFile), and JIT code generation (numba, warp).
Sanitization: Standard data science validation techniques are encouraged in the workflow (e.g., checking results against CPU baselines).

optimize-for-gpu