# AI Loop Library drop-in

Paste this block into a project `CLAUDE.md` (or save the sibling `SKILL.md` to
`.claude/skills/ai-loop-library/SKILL.md`). Short on purpose — it survives being read.

---

## Bounded work loops

For any repeated or "keep going until done" task, run it as a bounded loop, not open-ended
autonomy. Pick a matching loop from https://ailooplibrary.com/catalog.json (or the
`ai-loop-library` MCP tools if connected: `pick_loop_for_goal`, then `render_run_protocol`).

Rules, every loop, no exceptions:
1. Write state first: `docs/loops/<id>/progress.md` + `state.json` (goal, verification
   check, budget — default 8 rounds / 45 min / 3 consecutive failed verifications).
2. One change per round. New scope = new loop.
3. Same verification every round; record evidence before continuing.
4. Stop on: verifier pass, budget out, same failure after two different fixes, blocker,
   or approval boundary.
5. Approval boundary: money, production, outbound, deletion, accounts = never without a
   human. Public drafts and shared-branch PRs = pause and show first.
6. Finish with proof: rounds used, changes, final verification output, next human decision.