An autonomous research loop for important-but-stuck problems, operated by a scheduled Claude Code routine. The routine is the loop; this repo is the policy, the assets, and the memory.

• CLAUDE.md — the doctrine: one phase per run, decided by a state-machine table
• .claude/agents/ — planner (brief → spec), generator (one milestone → one artifact), evaluator (fresh-eyes pass/fail)
• .claude/skills/ — the asset library, extracted from successful manual sessions
• problems/<name>/ — problem.md (intake), spec.md (plan), state.md (memory/cursor), artifacts/, verdicts/

1. When nothing is active, the loop scouts a new stuck problem itself and writes the brief ([scout]). One problem at a time, in sequence. You can also add a brief by hand to problems/<name>/problem.md.
2. Next run: the planner writes spec.md with milestones and done-criteria, commits [spec]. No approval step. The loop continues on its own.
3. Runs alternate generate → evaluate, one phase per hour. PASS advances the milestone. FAIL retries once with the critique attached. A second FAIL blocks the problem and DMs you.
4. All milestones pass → the publisher drafts report.md and tweet.md, adds the problem to the table below, and DMs you. Publishing to the outside world stays human: you review and post.
5. The problem is published; the next wake scouts a fresh one. The loop never idles.

• Post the publications — when a DM says [published], review report.md and tweet.md, then post them.
• Unblock — when a commit says [blocked], read the verdict, fix the spec/rubric/state, push.
• Weekly 15-minute audit — the quality backstop now that no human reviews specs. Read two randomly chosen PASS artifacts. If one makes you wince, the fix goes into the evaluator rubric (versioned in evaluator.md). You're auditing the judge, not the work. Extract a SKILL.md when the log shows repeated freeform work.

Everything else runs without you. Watch the commit log: [blocked] needs you, [published] is your cue to post; the rest is the loop talking to itself.