Routines: Turn a Prompt Into a Schedule
The difference between a bot you ask and a bot you have is a schedule, a state file, and a quiet-day rule.
A prompt is an event. A routine is a system. The move from one to the other is small — a schedule, a state file, a report format — but it changes the economics completely: the bot starts producing while you sleep.
Anatomy of a routine that survives
One job per routine. “Morning digest” can watch five sources; it cannot also do weekly cleanup and social posting. When a second job appears, spawn a second routine — or a second bot (agent teams).
A named trigger. Vague: “check in sometimes.” Durable: “every day at 07:00 my time” or “every Monday morning.” Timezone it, or drift will find you.
A state file it updates. The routine reads last run’s leftovers and writes this run’s results:
seen.md — video ids I have already summarizedWithout state, every run re-reports the same things; with it, “new” actually means new.
A format that fits one screen. Routines die when their reports grow. Cap the format in the prompt: N items, M bullets, one link each.
A quiet-day rule. “Nothing today” is a feature, not a failure. Silence must be informative silence.
Choosing the cadence
| Cadence | Fits | Fails when |
|---|---|---|
| Hourly | prices, incidents, breaking news | sources update slower than you poll |
| Daily | digests, recaps, standups | report exceeds one screen |
| Weekly | competitor briefs, cleanup, reviews | anything urgent hides in it |
| On-event | ”when X posts” | trigger is vague or noisy |
Start daily. Hourly sounds impressive and mostly produces anxiety; weekly hides signal; daily at a fixed morning hour is the cadence humans actually sustain.
The maintenance sweep
Routines are living things. Once a month, one message:
Review your routine: list any config entries you haven’t used in 30 days, any rules you’ve had to interpret loosely, and anything in the log older than we need. Propose changes; don’t apply them.
Thirty seconds of review beats a silent failure you discover three weeks late. Pair it with a quarterly re-read of the prompt itself — prompt rot is real.
When routines multiply
Three routines reporting into one chat is a portfolio. Ten is a fleet, and fleets need what fleets always need: a roster, a foreman, and a shared report channel. At that point read Agent Teams and Track the fleet — and consider giving a bot the job of watching the bots.