Cuda

This is for writing, debugging, and optimizing CUDA kernels in Claude Code. It pushes a specific workflow: profile first with nsys or ncu before touching code, use printf in device code liberally when things break, and make small isolated changes instead of shotgun fixes. The compute-sanitizer patterns are solid for catching memory errors non-interactively, and the performance tables cut through the usual GPU optimization folklore. What's refreshing is the honesty about tooling limits: the skill explicitly says cuda-gdb often fails and sometimes you just need to stare at the diff between working and broken code. If you're doing GPU work and tired of guessing at bottlenecks, the measurement-driven approach here will save you time.

skills.sh installs: 256
GitHub stars: 231
last updated: Jun 3, 2026

View on GitHub

Install

$ npx -y skills add technillogue/ptx-isa-markdown --skill cuda --agent claude-code

Installs into .claude/skills of the current project.

Files

CUDA Programming Skill

Core Philosophy

Measure before guessing. GPU performance is deeply counterintuitive. Profile first, hypothesize second, change third, verify fourth.

Small, isolated changes. CUDA bugs compound. Make one change, test it, commit it. Resist the urge to "fix everything at once."

printf is your strongest tool. When debuggers fail, when tools produce inscrutable output, printf in device code reveals truth. Don't be embarrassed to use it extensively.

Sometimes, stare at the diff. Inscrutable segfaults are common. Tools often don't help. The human approach: minimize the diff, read it carefully, see the bug. This is legitimate and often faster than tooling.