The Agent Skills Directory

[COMMAND_EXECUTION]: The evaluation framework includes a shell script (evals/files/cudf-native-stream-handoff-boundary/code/run_smoke.sh) that invokes nvcc to compile a bundled CUDA C++ source file and executes the resulting binary. This is standard behavior for a GPU-focused technical skill to validate low-level synchronization logic and is restricted to local execution of bundled test files.
[PROMPT_INJECTION]: The skill provides instructions and examples for processing external data (CSV and Parquet) which constitutes a surface for indirect prompt injection.
Ingestion points: Multiple example scripts in the evals/ directory (e.g., etl_pipeline.py, parquet_pipeline.py) read from external files.
Boundary markers: None identified in the instructional code snippets.
Capability inventory: The skill demonstrates capabilities for file system operations and local compilation of GPU code.
Sanitization: Standard dataframe readers are used without custom sanitization of data contents, which is typical for this class of tool.
[EXTERNAL_DOWNLOADS]: The skill references official documentation and source code repositories for RAPIDS and NVIDIA (e.g., docs.rapids.ai, github.com/rapidsai/cudf). These references point to established, trusted repositories and are documented as standard resource links for the technology described.

accelerated-computing-cudf