# kernel-skills

A library of structured skill files that teach an AI coding agent how to write, optimize, debug and port high-performance compute kernels in CUDA, Triton and quantized workloads, rather than guessing at a good approach.

kernel-skills is a curated collection of skill files, each a structured engineering playbook an AI coding agent follows when writing, optimizing, debugging or porting a compute kernel. The project's premise is that, given only a vague prompt, an agent tends to skip gathering constraints, pick the wrong tiling strategy, ignore edge cases, and produce code that looks plausible but fails on real hardware; a well-authored skill instead forces it to gather constraints first, choose an appropriate algorithm and memory strategy, and reason explicitly about correctness.

The skills themselves are the product and can simply be read or pasted as markdown; a companion npm package wraps them in a versioned registry with a small command-line tool and a TypeScript API for scripted discovery. The project reports a measured example where a skill-guided softmax kernel stayed correct across shapes that made a naive version fail, including under adversarial inputs designed to overflow it.

It is explicitly not a kernel compiler, a benchmark harness, a model-serving runtime, or an autonomous agent itself; it never runs or measures the code it helps produce.

Website: https://github.com/tensormux/kernel-skills