← All articles

Build Your First Scheduled AI Routine: A Weekly Digest That Only Reads

October 5, 2026 · 3 min read

Most people meet AI agents as something they have to talk to. The more useful version runs while you aren't looking: every Monday at 7 a.m., a summary of your week is waiting, and you never typed a prompt. A weekly digest is the best first scheduled routine to build, because it's easy to set up, easy to judge, and, if you configure it right, it can't break anything.

Why a digest is the right starting point

A good first routine has three traits. The result is checkable, because you'll know in ten seconds whether the summary is accurate. A mistake is cheap, because a bad digest costs you a minute, while a bad auto-sent email can cost you a client. And it repeats, so it saves time every single week. A digest of your calendar and inbox hits all three, and it needs only read access: the agent looks, summarizes, and hands the result to you. It sends nothing and deletes nothing.

Step 1: Decide exactly what goes in

Vague routines produce vague output. Before you touch any settings, write down the sources and the question each one answers. For example: your calendar for the next seven days ("which meetings need preparation?"), and your inbox from the last week ("what is waiting on a reply from me, and what am I waiting on from others?"). Two sources is plenty for a first version. Resist the urge to add chat, documents, and task trackers on day one.

Step 2: Connect with the narrowest permissions

When your assistant asks to connect to your calendar and email, choose read-only wherever the option exists. If it only offers a broader grant, consider using a separate, limited account or label instead. The routine's job is to report, so there is no reason for it to hold the power to act. If you later want it to draft replies, add that deliberately, as its own change.

Step 3: Write the instruction like a brief for a new assistant

The prompt is where most routines succeed or fail. Spell out the output format, the order, and the limits. Something like this works well:

Every Monday at 7 a.m., review my calendar for the coming week and my inbox from the past seven days. Give me three sections: Meetings to prepare for (with one line on why), Replies I owe (oldest first, maximum ten), and Things I'm waiting on. If something is unclear, say so instead of guessing. Do not send, delete, or modify anything. Keep it under 300 words.

Notice the last two instructions. "Say so instead of guessing" gives the agent permission to admit uncertainty, and "do not modify anything" restates the boundary you already set in the permissions. Belt and suspenders is the right approach here.

Step 4: Schedule it, then audit the first three runs

Set the routine to run weekly and choose where the result lands: a notification, an email to yourself, or a saved note. Then treat the first three digests as a test, not a finished product. Compare each against your own sense of the week. Did it miss something that mattered? Did it flag noise? Adjust the instruction after each run: add a rule to ignore newsletters, or ask it to rank by urgency. By the third run, the digest usually settles into something you'd miss if it vanished.

Where to go from here

Once the digest earns your trust, you can widen it a notch at a time: add a source, let it draft (but not send) replies, or run a second routine for another recurring chore. Each step should be small, reversible, and reviewed by you. The pattern that makes agents worth having is simple: start with a job that only reads, check its work, and grant more only after it has earned it. This week, pick two sources, write the brief, and schedule your first one.

Sources: Practical guidance based on standard least-privilege permission practice; no statistics cited.