Most people use an AI assistant like a faster search box: ask a question, read the answer, close the tab. I use Muse, Meta's personal AI agent, as the operations layer for my family. It reads my email, runs my calendar, digs through three years of Amazon orders, and warns me before a credit-card annual fee hits.
Here's what actually works, after a few months of leaning on it hard.
1. Give it standing instructions, not one-off prompts
The single biggest change I made was to stop prompting and start instructing.
My son's school buries us in email: newsletters, sign-up deadlines, permission slips. I gave Muse one sentence: "Every action item in every email from my son's school gets recorded, reminded, and put on my calendar if it has a date."
That rule now runs itself. Every morning at 8:08 I get a short Chinese-language digest of my Gmail with a dedicated school section. Anything with a date, a survey due Friday, an after-school signup deadline, turns into a private calendar event with a reminder. I never re-prompt. The instruction persists, and the work just happens.
The tell: if you catch yourself asking for the same thing twice, that was never a prompt. It's a standing rule waiting to be set.
2. Hand it the boring multi-hour jobs
Nobody wants to page through 705 Amazon orders, so I didn't.
I needed every HSA-eligible receipt from the last three years. Muse went through my order history page by page, 2026 back to late 2023, flagged the eligible items (sunscreen, thermometers, bandages, reading glasses), downloaded 28 invoice PDFs, renamed each one YYYY-MM-DD-product, and handed me a zip.
My total time: one sentence.
The tell: the jobs AI is best at aren't clever, they're tedious. Anything you'd put off for six months because it's three hours of clicking is the perfect thing to hand over.
3. Connect your real accounts, then ask real questions
An assistant with no access to your data is a chatbot. One wired into your life is something else entirely.
I connected Gmail, Google Calendar, my iPhone's calendar, and my Amex cards through Plaid. Then I asked one plain question: "Look at my spending. What can I cut?"
It pulled two years of transactions and found about $5,200 a year in subscriptions I'd stopped thinking about: a $200/month API bill I assumed was bundled into something else, streaming I no longer watched, a kids' pizza-place membership left over from a birthday-party phase. (It also double-counted one at first, which is the reason for rule 5.)
Now every card's annual fee sits on my calendar with a reminder two weeks out, so the $695 Platinum renewal never ambushes me again.
The tell: "analyze my spending" only works if the assistant can see your spending. The connections are the difference between generic advice and answers about your actual money.
4. Let it work while you sleep
Two briefings land every morning at 8:08:
- Email digest: yesterday's Gmail boiled down to action items, in Chinese, school news always in there even when it's just FYI.
- Stock briefing: the previous close plus premarket news on trading days, with a weekend edition that covers the moves that mattered.
I didn't build a dashboard or open an app. The information just arrives, the same way the school deadlines just show up on my calendar. Ambient, not on-demand.
The tell: the best AI workflows are the ones you never open. Push, don't pull.
5. Keep a human in the loop for anything that leaves the house
The one rule I never bend: Muse doesn't send, book, buy, or post anything without my explicit yes. It drafts the email and shows me the exact recipients and body. It fills the cart and stops at the total. It proposes the calendar change and waits.
That's also what saved me when it read one subscription as two charges. Because nothing consequential happens without my sign-off, the mistake got caught at the review step instead of after the money moved.
The tell: delegate the work, keep the decisions. Autonomy on research, an approval gate on actions.
The shift
The mental model that changed everything: stop asking "what can I ask it?" and start asking "what would I hire an assistant to own?"
Email triage, receipt archaeology, a spending audit, deadline tracking, a morning briefing. Those are jobs, not questions. Hand over the job once, with your accounts connected and your approval rules clear, and it does it every day without you.
That's the whole post, really. Don't prompt harder. Delegate better.
Top comments (0)