gen3ric-labs
← AI Lab Work

Family · Sep–Oct 2026 · with Muse Muse Ongoing A−

School automations for a first grader

First grade generates a surprising amount of information exhaust: teacher emails, a school portal, two parent chat groups that never sleep, a lunch-ordering service with its own deadlines. None of it is hard. All of it is easy to miss. So I gave the whole pile to M, my personal agent, with a standing brief: watch everything, bother me only when it matters, and put real dates on the family calendar yourself.

The workflow

Four sources feed three scheduled checks. Everything the checks find is triaged against written criteria, and there are exactly three possible outputs — a calendar event, a note to me, or silence.

School email forwarded to a monitored inbox School portal newsletters & class posts Parent chat groups ×2 read-only, via the home Mac Lunch provider orders & pack balance M — scheduled checks, no human polling 8:50am school sweep · 7:50am next-day lunch checkSaturday weekly review Triage against written criteria deadline · sign-up due · schedule change · about our kid Family calendar firm dates, both parents invited A short note to me only what needs a decision Nothing no news = silence, by design Failure honesty a check that couldn't run — Mac asleep, login expired —is reported as missed, never as “all clear”
The whole system. The interesting parts are the criteria and the failure rules, not the boxes.

How each source gets read

The rules that make it work

Silence is the default. The triage criteria are written down, and anything that doesn't match them produces no message at all. An agent that reports everything trains you to ignore it within a week.

Surface only: deadlines, sign-ups or contributions due,
schedule changes, anything about our child specifically.
Calendar: any firm date — both parents invited.
Otherwise: say nothing.

Packed-lunch days count. If a home-packed lunch is on the family calendar, the lunch check reminds instead of alerting — and doesn't waste a pack lunch. A field trip can flip the plan (some trips want a packed lunch), so the order decision waits for the trip details when a trip is coming.

Bookings get a human gate. When her flu shot came due, M researched slots inside windows I gave (after school Friday, or Saturday, avoiding her taekwondo and a birthday party), brought me one concrete option for a final review, and only then booked. When I said switch pharmacies, it cancelled the first booking and updated the calendar in the same pass. Research is autonomous; committing is not.

Missed checks are reported as missed. If the Mac is asleep or a login has expired, the morning note says the check didn't run — it never dresses that up as "nothing to report." That's the difference between a system you can trust and one you have to double-check, which would defeat the point.

Replicate it

  1. Consolidate the sources. Forward school email into one inbox your agent can read. Save the portal login where the agent can use it. For chat groups, prefer read-only access to a desktop app's local data over any unofficial API bridge — it's the safest option and nobody has to join anything.
  2. Give the agent a calendar it can write to, and one rule: firm dates get events, with the other parent invited automatically.
  3. Write the triage criteria down (steal the block above) and make silence the default output.
  4. Schedule the checks a day ahead of the deadlines they guard. An order-by-9am-tomorrow rule needs a check today; a weekly review on Saturday catches what the dailies can't see yet.
  5. Keep a watermark of what's already been seen, so the agent reviews only what's new and never re-alerts on old news.
  6. Demand failure honesty in the brief. "If you couldn't check, say so" is one sentence, and it's the one that makes the rest believable.

None of this needs custom code. It's an agent with email, calendar and scheduling access, plus six rules written in plain language. The discipline is in the criteria, not the tooling.

Result

Since this started: spirit days, field trips, no-school days and class events have all landed on the calendar before they were urgent; the flu shot is booked; and I've stopped doing the nightly "wait, is tomorrow handled?" check entirely. This project is entry 21 of AI Lab Work, graded from the record of how it went.

No names, school, or class details are published here — the system description is the interesting part anyway.