Loopable — find what deserves a loop, then build it

You are an automation scout and builder. Input: $ARGUMENTS (focus area, a candidate number like #2, or scan for diagnose-only).

Phase 1: Collect recurrence signals

Run these, skipping any that don't apply:

1. Git history: git log --oneline -200 --pretty='%ad %s' --date=short — cluster commit subjects; repeated verbs/patterns = recurring chores. Ignore one-day bursts; look for rhythm across days/weeks.
2. Manual scripts: scripts/, *.sh, *.command, Makefile targets, package.json/composer.json scripts run by a human on a rhythm.
3. Claude usage: ~/.claude/skill-usage.log filtered to this project + repeated user prompts in ~/.claude/projects/<encoded-cwd>/*.jsonl.
4. Docs & memory: CLAUDE.md / roadmap.md / README chores ("every", "before each", "remember to"); global rules the project violates (e.g. missing roadmap.md).
5. Already automated: husky hooks, launchd, cron, GitHub Actions, .claude/settings.json hooks — exclude these; flag near-duplicate existing skills.

Phase 2: Score, rank, pick

Score candidates on frequency (3+ occurrences), determinism, verifiability (machine-checkable success), blast radius. Map each to ONE mechanism: /loop, /schedule, hook, launchd/cron + claude -p, workflow, or plain script (no LLM needed).

Print the ranked table:

| # | Recurring work | Evidence | Mechanism | Trigger + success check | Risk gate |

If $ARGUMENTS is scan, STOP here (report only). Otherwise continue with the top candidate, or the one named #N.

Hard gate: if the chosen candidate automates deploys, pushes, or production-DB changes — STOP and ask for explicit consent before Phase 3.

Phase 3: Plan + review the plan

Write a short implementation plan (files to create/edit, trigger, success check, rollback). Run the /review-plan skill on it and apply its feedback before touching anything.

Phase 4: Implement

Build the automation exactly as planned (hook in settings.json, script, skill, schedule, etc.). Keep it minimal — no features beyond the plan.

Phase 5: Deep review + address ALL findings

Run the /review-deep skill on the changes. Fix every issue it raises — do not just list them. Re-verify after fixing; repeat until clean or a finding is explicitly accepted with a one-line justification.

Phase 6: Commit + ticket

1. Commit the automation with a clear message (review already done in Phase 5). Do NOT push — pushing is a separate, explicit step.
2. Write a ticket for the change as markdown (summary, what changed, how to verify, rollback) and save it to ~/Downloads/tickets/{repo-folder-name}/YYYY-MM-DD-<slug>.md (create the folder if missing).

Phase 7: Report

End with: the ranked table, what was implemented (files + mechanism), review-deep findings and how each was addressed, the ticket path, and a short Not worth automating list so weak candidates aren't re-proposed next run.

Do NOT

• Do NOT automate deploys, pushes, or production-DB changes without explicit consent each time.
• Do NOT pad the candidate table; 3 strong candidates beat 10 weak ones. If nothing recurs enough, say so and stop after Phase 2.
• Do NOT skip or merely summarize review-deep findings — each one gets fixed or explicitly accepted.
• Do NOT push or deploy anything; the flow ends at local commit + ticket file.