An agent harness that compiles a model into one provably-correct, self-retargeting CUDA megakernel and self-tunes it past cuBLAS at batch-1 LLM decode, paper: https://arxiv.org/abs/2606.09682
6 files for Claude Code, Codex and OpenCode: amk-megakernel-optimizer, amk-autoresearch, amk-compile, amk-optimize and 2 more — 2,279 tokens loaded in every session.
AGENTS.md B 2,279 tok .claude/skills/megakernel-optimization/SKILL.md A 57 tok .claude/commands/amk-autoresearch.md A 25 tok .claude/commands/amk-compile.md A 29 tok .claude/commands/amk-optimize.md A 27 tok .claude/agents/amk-megakernel-optimizer.md A 62 tok These files are RightNow-AI/AutoMegaKernel's own configuration — they tell Claude Code, Codex and OpenCode how to work on this repository, so they are not mods to install elsewhere. Copy one as a starting point and replace the parts that are about this project.