01 AutoResearch-RL 2026 GitHub ↗ An AI agent that runs RL post-training experiments overnight — edits the config, trains for ten minutes, keeps or discards, and moves to the next idea. Built on prime-rl and verifiers. 02 Synthetic Self-Improvement 2026 GitHub ↗ A Claude Code skill where Claude designs the synthetic data, environment, and rewards to post-train a smaller model — in a loop, until a wall-clock budget runs out. 03 RL Environment Audit 2026 GitHub ↗ Audits verifiers RL environments from the Prime Intellect Hub before you train on them — six checks and a scorecard out of ten, because a broken reward silently teaches the policy garbage. 04 Flash Attention From Scratch 2025 GitHub ↗ CUDA implementations of Flash Attention 1 and 2 written from scratch, following the original papers — tiling, fused online softmax, and FP16 and pipelined variants.