CUDA Kernels as Agent Skills

Hugging Face just added something interesting to their kernels repo: agent skills for CUDA kernel generation.

Now, instead of writing CUDA kernels yourself, you give a structured skill file to Claude or Codex, and the agent writes them for you.

The skill is ~550 tokens, pointing to other docs from there. It contains GPU architecture specs, memory optimization patterns, PyTorch bindings templates, and benchmarking scripts. The agent reads it, and when you ask for a kernel, it generates a complete project adapted to your target GPU such as H100, A100, or T4, each with its own compute capability, memory bandwidth, and shared memory config. Ready to compile and test.

According to their benchmarks: 1.88-2.47x speedups on LTX-Video and Qwen3-8B.

What caught my attention isn't the CUDA part. It's the approach.

They took years of GPU optimization expertise and compressed it into a skill file that any coding agent can consume. Not a tutorial. Not documentation. A structured prompt that turns a generalist agent into a specialist.

And the skill approach itself is worth stealing. If you've got domain expertise that's hard to transfer, package it as an agent skill. A bunch of structured tokens of guidance beats 50 pages of docs that nobody reads.