---

name: self-improving-skill-overnight-loop description: Run Karpathy's auto-research loop on a Claude skill: try a tweak, run binary assertions, keep or revert based on pass rate, and iterate all night.

Self-Improving Skill Overnight Loop

What This Does

The creator adapts Karpathy's three-file loop to skills by having Claude read the skill.md, propose a change, run five test prompts, and score outputs against 25 true or false assertions. If all pass, git commit; if any fail, get reset and try again, looping without asking for permission until interrupted or a perfect score is hit.

What You Get

• evals.json with true or false binary assertions
• auto-modify skill.md loop
• git commit or git reset per iteration
• per-iteration pass rate log
• overnight autonomous run

Customize

• assertions: array of binary tests per test prompt
• test_prompts: sample inputs to grade
• max_iterations: cap to prevent infinite loops
• commit_on_pass: whether to auto-commit each improvement
• stop_when: perfect score, manual, or time limit

Sample Output

First run scores 23 of 24 assertions, Claude adds one clarifying rule to the skill.md, second run hits a perfect score and commits.

Target Audience

Skill authors iterating on production skills

Source

https://youtube.com/watch?v=wQ0duoTeAAU