Claude Code’s Overnight Shift: Iterate, Verify, Repeat

What it does Autoresearch is a set of prompt-driven commands for Claude Code, OpenCode, and OpenAI Codex that automates a tight feedback loop: make one atomic change, run a mechanical verification, commit the result, and revert it if the metric worsens. It generalizes Andrej Karpathy’s autoresearch script—originally built for overnight ML experiments—to any domain where progress can be measured with a number. Before each iteration the agent reads git log and a TSV results file, so it learns from near-misses without human hand-holding.

The interesting bit The architecture is deliberately thin: a 41-line routing file dispatches to twelve self-contained command files, cutting per-invocation tokens by 95% compared to the previous monolithic 813-line skill file. That frugality matters because it leaves more context window for your actual codebase. The safety hooks are also notably paranoid—blocking rm -rf, credential reads, and force-pushes by default, and warning when a file balloons past 400 lines.

Key highlights

• Thirteen specialized commands, from core iteration (/autoresearch) to security audits (/autoresearch:security) and adversarial reasoning (/autoresearch:reason).
• Bounded by default: every command ships with a sane iteration cap; unlimited looping requires an explicit opt-in.
• Git-as-memory: each experiment is committed with an experiment: prefix, and the agent reviews git diff before picking its next move.
• Nine safety hooks that run on every session, including context compaction reminders and a simplify-gate that blocks shipping files over 800 LOC.
• Supports Claude Code, OpenCode (underscore syntax), and OpenAI Codex ($autoresearch mention syntax).

Verdict Worth a look if you live in Claude Code or Codex and have a tedious, measurable task—like raising test coverage or shrinking bundle size—that benefits from brute-force iteration. If you expect a standalone CLI tool or magic without a defined metric, this is just a very organized pile of prompts.