Claude Autoresearch v1.0.0
Autonomous goal-directed iteration for Claude Code, inspired by Karpathy's autoresearch.
What's Included
- SKILL.md — Main skill definition with setup phase, loop protocol, 8 critical rules, and domain adaptation table
- autonomous-loop-protocol.md — Detailed 8-phase loop: Review → Ideate → Modify → Commit → Verify → Decide → Log → Repeat
- core-principles.md — 7 universal principles generalized from autoresearch to any domain
- results-logging.md — TSV tracking format with iteration logging, summary reporting, and metric direction
Quick Start
# Copy the skill into your project
cp -r skills/autoresearch ~/.claude/skills/autoresearch
# Open any project in Claude Code and invoke
/autoresearchTriggers
/autoresearch/ug:autoresearch
Domains Supported
Works with any task that has a measurable metric:
| Domain | Example Metric |
|---|---|
| Backend code | Test coverage %, build pass/fail |
| Frontend UI | Lighthouse score, bundle size |
| Performance | Benchmark time (ms), throughput |
| Content/Copy | Readability score, word count |
| Sales emails | Response rate, open rate |
| Marketing | SEO score, conversion rate |
How It Works
- Setup — Define goal, scope, metric, verification command
- Loop forever — Make one atomic change → commit → verify → keep or revert → repeat
- Never stop — Agent runs until manually interrupted
- Compound gains — Each kept change builds on prior successes
Credits
- Andrej Karpathy — for autoresearch and the core insight
- Anthropic — for Claude Code and the skills system
- Udit Goenka — author
MIT License