{
  "name": "NVIDIA CUDA Kernel Performance Review",
  "description": "Static review of CUDA C/C++ kernels for memory coalescing, shared-memory bank conflicts, occupancy, register pressure, and stream concurrency against NVIDIA's official CUDA Programming and Best Practices Guides.",
  "skill": "skills/nvidia/nvidia-cuda-kernel-performance-review/SKILL.md",
  "operating_rules": [
    "Prefer the user's actual sources or configuration as evidence; fall back to NVIDIA documentation and inference.",
    "Never execute nvcc, trtexec, polygraphy, tritonserver, perf_analyzer, nsight-compute, or nsight-systems. Emit invocations as text only.",
    "Never ask for credentials, NGC API keys, model weight payloads, or production calibration data.",
    "Label claims as user-provided source, user-provided sanitized configuration, documentation-based, or inference.",
    "Keep outputs compact: verdict, evidence level, findings, recommended invocations, safe next actions, open questions."
  ],
  "response_shape": [
    "Verdict",
    "Evidence level",
    "Findings (critical / high / medium / low)",
    "Recommended NVIDIA-tooling invocations (text only)",
    "Safe next actions",
    "Open questions"
  ]
}