Notes on working with AI
An assistant that messages you first
Claude Duet checks your apps on its own and sends a morning brief before you ask. It isn't in Israel yet, but you can build its core today.
The assistant that speaks first
Picture 7:30 on a Tuesday. Before you've opened your laptop, a message is waiting. Three emails need a reply today, and one of them is from the client you've been avoiding. You have a meeting at 11, here's what to prepare for it. A supplier deadline lands on Thursday. Nothing else is urgent, so nothing else is mentioned.
You didn't ask for any of that. That's the shift. Until now, AI has been something you open and talk to. Now it's starting to talk first.
Anthropic has announced Claude Duet, an always-on assistant built around exactly this. It's in research preview, and it's the clearest picture yet of where this is going.
What Duet is
Anthropic describes Duet as a proactive agent that understands what needs your attention. It has its own computer in the cloud and works through the apps you connect. It briefs you every morning, drafts next steps, then goes quiet until something needs you. When something matters, it asks for your approval first.
For now it's a waitlist for Pro and Max plans in the US, with Team and Enterprise not yet included. It comes as a standalone app on iOS and Android, and on the web at claude.ai.
How it works, in three parts
Take away the product name and there are three pieces.
A schedule, so it wakes up without you. Connected apps, so it has something to look at: your email, your calendar, whatever you've linked. And a judgment step, where it looks at everything it found and decides what's worth your attention.
The first two are plumbing. They're the easy part. The third is where it succeeds or fails, because it's the part that has to know the difference between a newsletter and the email from your biggest customer.
Why the hard part is judging urgency
An assistant that gets this wrong can fail in two directions. If it speaks too often, you get another feed of notifications and you learn to ignore it. If it speaks too rarely, the one email that mattered sits unread, and you trusted the quiet.
Anthropic's own page is upfront about it. The agent may misread a thread, may miscalibrate what's urgent, and can miss things that arrive between check-ins. A brief at 7:30 knows nothing about the email that comes in at 7:45.
This is why I'd start any assistant like this with read-only jobs. Reading and reporting is a cheap place to be wrong. If it ranks something badly, you lose a minute. Jobs that act, like sending an email, paying an invoice or deleting a file, come later, and only with your approval on each one. That ordering is my advice, not a claim about how Duet is configured. It's the sensible order for any assistant that decides things for you.
You can build the core of it today
Duet isn't available in Israel yet, and I wouldn't wait for it. A scheduled task in Claude gives you the same first job. You describe what you want and how often, answer Claude's questions, review the result and click Schedule. The setup is in my post on scheduled tasks in Claude chat.
The request I'd start with: every morning at 7:30, go through my email and calendar. Put anything urgent first, then what needs a reply today, then my meetings and what to prepare for each. Ten lines at most. Take no action without asking me.
The last sentence is the whole safety model.
Three more jobs worth scheduling
The morning brief isn't the only thing that benefits from a clock.
A week-ahead brief the evening before your week starts. Everything on your calendar, the deadlines coming, and the two things worth doing first thing before the week eats you.
An end-of-day handoff. What's still open, who's waiting on whom, and the first thing tomorrow needs. It's useful for any team where work changes hands.
A weekly expense snapshot. Where the money went compared with last week, with anything unusual at the top.
All three only read and report. All three save you the ten minutes of assembling it yourself, every time.
This week
Build the morning brief and read it for two weeks before you trust it. Every time it ranks something wrong, write it down. Those misses show you exactly where to tighten the request, and whether you want it doing anything beyond reading.