Custom Kernels for All from Codex and Claude

This post introduces an agent skill that enables coding agents (Claude and Codex) to write production-ready CUDA kernels for HuggingFace's diffusers and transformers libraries.

• •The skill (~550 tokens of structured guidance) packages GPU architecture-specific knowledge for H100, A100, and T4, covering memory access patterns, vectorization strategies, and PyTorch bindings
• •Agents install the skill via `kernels skills add cuda-kernels --claude` and generate complete kernel projects including CUDA source, PyTorch C++ bindings, build config, and benchmark scripts
• •For LTX-Video (diffusers), RMSNorm kernels achieved an average 1.88x isolated speedup and 1.43x end-to-end speedup on H100 when combined with torch.compile
• •For Qwen3-8B (transformers), RMSNorm kernels achieved an average 1.94x isolated speedup, scaling from 1.58x at 128 tokens to 2.47x at 8192 tokens