Day 6 of 16 · Friday · Learning

/loop and ScheduleWakeup vs fragile LaunchAgentsCapability

Two tools let Claude pace its own recurring work while you watch — /loop (repeat a prompt on an interval) and ScheduleWakeup (let the model decide when to wake up next). Unlike your 06:15 LaunchAgents, every iteration happens inside a live session, so when it breaks you SEE it break.

Catch-up progress

6/16

Why this matters to you

Your whole morning system runs on LaunchAgents — com.tom.pdb-plan-day, com.tom.pdb-prep-big3, com.tom.pdb-language-lesson, app.zorc.learning-lesson. They're the right tool for 06:15 when you're asleep. But they fail silently: the RemoteTrigger that emailed your language lesson died twice (5/27 and 5/28) and last_fired_at still said 'success' because the trigger fired even though the work didn't. /loop is the opposite trade — it runs while you're present, so a failure lands in the conversation instead of a log nobody reads.

A LaunchAgent is macOS's cron: a plist in ~/Library/LaunchAgents tells the OS 'run this command at this time' and walks away. It's unattended and durable — perfect for the 06:15 plan-day run — but the OS only guarantees it STARTED the process. Whether the work actually succeeded is invisible unless you build verification in (which is exactly why your CLAUDE.md automation-hygiene rules exist).

/loop is a different shape: you're in a Claude session and you say '/loop 10m /some-command'. Claude runs it now, waits 10 minutes, runs it again — inside the same conversation. Every iteration's output is right there. If iteration 3 errors, you read the error. It's for work you want to babysit: polling a deploy, watching a file land, re-checking a flaky thing until it's green.

ScheduleWakeup is the self-pacing version. Instead of a fixed interval, Claude looks at what it's waiting for and picks its own next wake-up time — 'CI takes ~8 min, I'll check back in 270 seconds' vs 'nothing to do, sleep 20 minutes'. You hand over the cadence decision to the model. The mental model: LaunchAgent = fire-and-forget alarm clock; /loop = a kitchen timer you keep resetting while you cook; ScheduleWakeup = telling a smart assistant 'wake me when it makes sense'.

Worked example

You don't need to write any code to try /loop — it's a slash command you type in a session. Here's the smallest useful one against your own output repo:

/loop 5m check whether output/plan-day/today.html has been updated in the last hour; if not, tell me, otherwise just say 'fresh' and keep waiting

'/loop 5m' = repeat the rest of the line every 5 minutes, in this session.
The prompt is plain English — Claude re-runs the check each tick (here, looking at the file's mod time in your PDB output repo).
Because it's in-session, the moment the check fails you get a sentence in chat — no plist, no log file, no silent last_fired_at.
To stop it, you just tell Claude to stop, or close the loop — there's no launchctl unload dance.

▶ Do it now

Pick ONE thing you currently check by hand-refreshing — pdb.zorc.app loading, a git push landing, the language email arriving in your inbox.
Start a /loop on it with a sane interval (5m for something that changes in minutes; don't go under 60s).
Let it run two or three iterations, then watch what happens when you deliberately make it fail — e.g. point it at a file that doesn't exist — and confirm the failure shows up in the conversation immediately.
Then stop the loop. The lesson isn't the loop itself — it's feeling the difference between a failure you SEE and the silent 5/27 kind.

Gotchas

/loop and ScheduleWakeup only run while a session is alive — they are NOT a replacement for your 06:15 LaunchAgents. The minute you close the session, the loop is gone. Critical daily jobs stay on launchd.
Don't set a tiny interval to 'poll' for background work the harness already tracks — when a tracked job finishes, Claude gets re-invoked automatically. Polling every 60s just burns tokens.
Intervals under 5 minutes keep the prompt cache warm; crossing 5 minutes pays a cache-miss cost. For idle waiting, a longer 20-30 min tick is cheaper than hammering every minute.

Go deeper: Your automation-hygiene rules (why silent failure is the enemy)

One-card takeaway

Use launchd for work that must happen while you sleep; use /loop and ScheduleWakeup for work you want to watch happen — the difference is whether a failure surfaces in your face or rots in a log.

← Prev Index Next →