Traction scoreGitHub stars can be faked, so popularity alone can be misleading. Traction Score looks for broader signs of recent attention, adoption, and active maintenance.
This skill enforces a disciplined workflow for CUDA kernel work: profile before guessing, isolate one change at a time, and use printf, compute-sanitizer, and nsys as primary tools. It is built for GPU compute engineers, HPC performance engineers, and debugging specialists who need a repeatable path from bug to fix and from bottleneck to speedup.
WHO IT'S FOR
GPU Kernel Developers
debugging and optimizing CUDA kernels
AI/ML Engineers
accelerating deep learning models with CUDA
HPC Researchers
writing and profiling CUDA simulations
Debugging and Tooling Engineers
troubleshooting GPU memory errors and race conditions
Compatible AgentsThe repository documents support for these agents. The skills may also work with other agents that can load SKILL.md files, but they may need some setup or small changes.
Compatibility not documented
Use this skill
Follow the documented setup, then try a first task.
Path to the CUDA source project, program, or executable
Small failing input or reproduction steps
Expected behavior
Actual error or observed behavior
Check the repository for this skill’s setup.
Use the cuda skill to perform a first debugging pass on my CUDA program. Reproduce the failure with the smallest practical input, inspect the kernel for likely causes, and recommend one isolated next change. If execution is available, begin with device-side printf tracing and then run the appropriate compute-sanitizer check. Program or executable path: <path-to-program>. Failing input or reproduction steps: <failing-input-or-steps>. Expected behavior: <expected-behavior>. Actual behavior: <actual-behavior>.
