Autoresearch

Routes you through Karpathy's autonomous ML search loop: write program.md with your research goal, let the agent edit train.py, keep runs that lower val_bpb in 300 seconds, revert everything else. The skill enforces the immutable harness rule (prepare.py and eval stay frozen once your session starts) and picks the right mode whether you're doing first-time setup, refining your research charter, running the overnight loop, or interpreting results.tsv. It draws a hard line between this and prompt eval tooling like LangSmith or Braintrust. Works best when you have a real training repo, a GPU with ~40GB VRAM, and want disciplined ratcheting instead of hero rewrites.

skills.sh installs: 177
GitHub stars: 22
last updated: Jun 3, 2026

View on GitHub

Install

$ npx -y skills add akillness/oh-my-skills --skill autoresearch --agent claude-code

Installs into .claude/skills of the current project.

Files

autoresearch

Autoresearch is a closed-loop ML experimentation workflow:

human writes program.md
agent edits train.py
prepare.py stays fixed
every run gets the same 300-second budget
lower val_bpb wins
regressions get reverted

This skill should behave like a routing-first front door, not a giant tutorial. Pick the user's mode, enforce the immutable-harness rules, then hand them to the smallest useful script or reference.