Six Claude Routines Quietly Maintaining My Repos
I pointed Claude Desktop routines at my two vibe-replay folders and let them run on cron. Code-quality sweeps every morning, deps and docs every week — the boring chores I always forget.
I have two working copies of vibe-replay on my laptop — vibe-replay/ for the main repo and vibe-replay2/ for in-flight experiments. Both accumulate the same kind of low-priority maintenance debt: stale deps, drifting README, plans I wrote two weeks ago that nobody pruned.
So I stopped pretending I’d ever do those chores by hand and handed them to Claude routines.
Six cron-scheduled agent runs, each scoped to one of the two folders:
- Daily code quality — every morning at 7 AM. A short sweep for obvious smells, dead code, type hole-punching. Anything actionable becomes a draft PR.
- Weekly deps update — Mondays at 2 AM. Bumps non-major versions, runs the test suite, opens a PR if green.
- Weekly upstream schema watch — Tuesdays. Checks Claude Code’s JSONL schema and Cursor’s session format for changes vibe-replay’s parsers need to track.
- Weekly readme docs sync — Wednesdays. Diffs the README against the actual CLI surface and flags drift.
- External research routine — Saturdays. Scans for new posts and projects in the AI-session-tooling space worth a closer look.
- Weekly plan status sync — Sundays. Walks the planning docs and marks anything that’s been shipped or abandoned.
The whole thing took about ten minutes to set up. Each routine is just a markdown skill file plus a cron expression, scoped to one folder so it can’t wander.
The honest result: maybe one in three runs produces something I act on. The other two are no-ops, which is fine — that’s the point. The cost of a no-op is near zero; the cost of me forgetting to bump a dep for three months is a Friday night debugging a CVE.
If you’re curious how the underlying machinery works — Claude Desktop’s autonomous agent mode writes its transcripts somewhere completely different from regular Claude Code sessions, which is why getting vibe-replay to see these runs took some work. I wrote that up separately in Capturing Claude’s Autonomous Agent Mode. And if you’ve never seen vibe-replay itself, the origin story is the shortest intro.
For now I’m letting the routines run for a couple of weeks before I touch the configuration again. Half the value of automation is resisting the urge to fiddle with it.