Your first scheduled task

The version of this that goes wrong looks like this. A job you already do by hand every week, and a vague intention to automate it. Then a prompt written inside a chat that already knows which account and which week you mean. You put it on a schedule and it fires overnight. What comes back is thinner than the version you ran yourself. Nothing errored. The task simply did not know what you knew, and you find that out after trusting it once.

The hand version costs you the front of every working day. It pays nothing back. The scheduled version costs one prompt, one interval, and one place to switch it off. That trade is not close.

Agents handle my inbox, my reporting and my ops. Not one of them started as a clever build.

A scheduled task is not a reminder, not a saved prompt, not a chat you left open. It is a stranger with your keys who turns up at 07:00 whether you are there or not.

The rules

Schedule the job you already do, not the job you wish you did. Your first one should be dull. Something you have run by hand ten times. You already know what a good answer looks like, and you spot a bad one in three seconds. If you cannot grade the output, you cannot schedule the work.

A scheduled run inherits nothing. No chat above it, no memory of the session where you wrote it, no you. Everything it needs goes in the prompt. The paths, the names, the shape of the reply. Step 4 takes that apart properly.

Your first task reads and does not write. Reads are free and they undo themselves, and a write that goes wrong at 03:00 costs you the morning. The full case for going read-only first is in give-your-ai-the-keys step one. Earn the writes here later.

Know where the off switch lives before the task exists. Not the prompt, the enabled flag, wherever your scheduler keeps it. A prompt that says the task is finished is not an off switch. Step 3 takes that apart.

Do it now

Pick one read-only job you already run weekly. Write it out cold:

TASK: read <the thing>, report <the numbers>.
WHERE: <exact path, exact account, exact table or folder>.
RETURN: <the shape — three bullets, a table, a one-line status>.
IF THE READ FAILS: say so. Do not report clean.

Set the interval in UTC and convert your local time yourself. Run it by hand three times and log the result of each one, before the schedule ever touches it. Then write down where its enabled flag lives, in the same place you keep the task.

A worked example

Say you run an online candle shop, just the two of you. Ellie packs the orders. Sam pours the candles. Every morning one of them opens the shop admin and the stock sheet, works out what to pack and which scents are running low, and loses the first half hour of the day to it.

They have done it by hand a hundred times. A wrong answer would jump out. That makes it the right first task.

TASK: read yesterday's paid orders and the stock sheet, report what to pack and what is low.
WHERE: shop admin, Orders, status Paid, yesterday 00:00 to 23:59 UK time.
       Stock sheet, tab Finished, jar counts in column D.
RETURN: three bullets. Orders to pack. Scents under 10 jars. Anything odd.
IF THE READ FAILS: say so. Do not report clean.
Runs:        daily at 06:00 UTC, which is 07:00 in the UK until the clocks go back
Hand runs:   Monday, Tuesday, Wednesday. All three logged, all three bullets
Off switch:  enabled flag in the scheduler's task list, noted in this file
First run:   Thursday 06:00 UTC. Three bullets, the same shape as the hand runs

One task, one interval, one switch. The rest of the track builds on it.

The failure you will hit

The first scheduled run will be worse than your test run. Your test ran inside a conversation that already knew things. Which account, which week, which report. The schedule knows none of it. The second most common failure is timing. It fires an hour out in a UK summer, more elsewhere, because you set local time on a scheduler that thinks in UTC. Do not read the description to check. Read the timestamp.

Not for you if

You have never done this job by hand, so you cannot tell a good output from a bad one. Automate something you understand, or you are paying a model to generate text on a timer.

The Vault is the paid library, and Operations is one of its categories. The price is on the page.