claude-autoresearch-skill

Autonomously research an ML task and run MANY bounded experiments to find the best config — a fixed-budget edit→train→eval→keep-or-discard loop in the spirit of karpathy/autoresearch, wrapped in the ml-intern orchestrator model and fanned out with a Claude Code dynamic workflow. Runs LONG: an iterative generational loop (mims-harvard/AutoScientists style) where parallel agent teams propose hypotheses, peer-critique them before spending any GPU, share findings on a common board, promote a champion, and keep going until budget/stagnation/convergence. Triggers when the user wants to "run many experiments", "sweep / search for the best config", "beat a benchmark", "do an ablation", "autoresearch X", "run for a long time / overnight / for days", or "find what improves metric Y on dataset Z". Deep-researches existing solutions across the internet FIRST (fan-out web search + PapersWithCode + GitHub, sources cross-checked into a cited DEEPRESEARCH.md), then ASKS where to get GPUs ("cards") and data before spending any compute, generates an experiment matrix seeded from diverse literature angles, runs it as a background workflow under an explicit budget, keeps a running leaderboard + shared findings board, verifies winners, and reports the best config. Reuses ml-intern's notify.sh + hf_push.sh for milestone alerts and HF Hub publishing.