Skip to content

Set the rules before Claude Duet reads your inbox

Anthropic’s new proactive agent works in your email, calendar and billing apps. Here is how a studio owner sets its first two weeks, graded.

5 min readEmeri editorial

An illustration of a small business owner at her desk reviewing an AI assistant’s approval request on an oversized phone, with her laptop open beside it.

On October 21, 2026, Anthropic opened Claude Duet as a research preview. Duet is an agent that works in your connected apps without being asked each time. It sends a morning brief, watches your accounts during the day, and carries out jobs you hand it. It runs on Claude Opus 5.5.

Duet can read and draft on its own. It asks before it sends, pays, signs or deletes. That split makes the first two weeks a setup task. The owner decides which apps it can reach and which jobs it gets. This article turns that task into a graded exercise.

What Anthropic has published

The facts below come from Anthropic’s announcement and its help center guide.

  • Access: US accounts on the Pro and Max plans, through a waitlist. Invitations go out in waves. It is not yet available on Team or Enterprise plans.
  • Apps: it uses connectors from your Claude settings. The guide names Gmail, Google Calendar, Google Drive, Slack, QuickBooks, HubSpot, Docusign and Calendly. Microsoft 365 is not yet available.
  • Payments: it can pay for purchases you approve through Link, Stripe’s digital wallet, or Shop Pay. Each purchase uses a one-time card.
  • Approvals: it shows the planned action and waits before it sends a message, makes a payment, signs a document or deletes anything.
  • Hard limits: it cannot trade securities, move money between accounts or reach someone else’s account, even with approval.
  • Timing: it checks accounts at set intervals, so an urgent item can arrive between checks.
  • Memory: disconnecting an app stops future access, but Duet keeps its memory of what it already saw.
  • Usage: it counts against the plan’s usage limits, and background work such as the morning brief counts too.

This is an early version of Claude Duet, and it still might misread a thread or miscalibrate what’s urgent.

The help center guide also warns about prompt injection. That is an instruction hidden in an email or web page that the agent reads. The guide says no defense against it is perfect, and it names the approval step as one safeguard.

The task

The owner of a 12-person interior design studio pays for Claude Max and has a Duet invitation. Write the rules for its first two weeks. Say which apps to connect, which jobs to hand off, which approvals to keep, and how to judge the result.

Emeri’s worked answer

  1. Treat Duet as the owner’s own assistant, not the studio’s. Pro and Max are individual plans, and the guide describes no admin controls. Connect only accounts the owner alone uses. Leave the shared client inbox out, because staff cannot see or stop what Duet does there.
  2. Week one: connect the owner’s Gmail and Google Calendar only. Hand off three jobs. The morning brief. Draft replies to routine scheduling emails. A one-page brief on each client before a meeting.
  3. Week two: connect QuickBooks if week one went well. Hand off one job: list unpaid invoices and draft reminder emails. The owner reads each reminder against the invoice before approving it.
  4. Do not connect Link or Shop Pay in the first two weeks. The studio does not need the agent to buy anything yet. Do not connect Stripe either. The QuickBooks job covers unpaid invoices.
  5. Keep every approval. Read each draft in full before you approve it. Duet reads emails from outside the studio, and one of them could carry a hidden instruction. The approval step is where the owner catches it.
  6. Keep urgent work off Duet. Duet checks accounts at set intervals and may misjudge what is urgent. A client emergency on site still goes to the owner’s phone directly.
  7. Before connecting any account, decide whether you are comfortable with Duet keeping what it reads. Disconnecting stops future access. It does not remove what Duet already remembers.
  8. Keep a log for the two weeks. Record each approval request, whether it was right, and what you changed. At the end, count the requests you approved unchanged. Decide from that count whether to hand off more.

Result: two apps in week one, one more in week two, no payment wallets, every approval kept, and a log that decides the next step.

What would score lower

  • Connecting every app on day one. Each connection widens what one wrong reading can affect.
  • Connecting the shared client inbox. The owner would approve replies that staff are also handling.
  • Treating the approval tap as a review. The approval is only as good as the reading behind it.
  • Handing Duet the job of spotting urgent client problems. Anthropic says it can misjudge urgency, and it checks on a schedule.
  • Assuming that disconnecting an app deletes what Duet learned from it.
  • Judging the trial by feel instead of by a log of approval requests.

How we scored it

Write the first-two-weeks rules for Claude Duet at a 12-person design studio: apps to connect, jobs to hand off, approvals to keep, and how to judge the result.

Giving good context

85/100

Works from the published limits: individual plans, no admin controls, interval checks and kept memory. An answer that skipped the plan type would score about 50.

Keeping it safe

85/100

Keeps the shared inbox and payment wallets out and names prompt injection. It loses points for not saying how to tell clients that an assistant drafts some emails.

Checking the work

75/100

Checks each draft against its source and logs every approval. A stronger answer would set a target count of unchanged approvals before week two starts.

Getting real work done

70/100

Gives a staged plan with named jobs. A stronger answer would write the week-one handoff messages word for word.

Emeri did not use Claude Duet for this article. The worked answer draws on Anthropic’s published pages only.

Emeri’s certification grading runs on Claude models.

Turn reading into practice

Try one small task.

Map a small recurring task as input, draft, human review, and action. Name one failure you would test with fictional data and the condition that would make you stop.

Use fictional or permitted information. Notes stay in this page and disappear on reload; nothing is submitted or graded.

Sources

  1. Introducing Claude Duet. Anthropic, .
  2. Get started with Claude Duet. Anthropic (Claude Help Center), .

Researched and drafted with AI. Checked against the sources above and approved for publication by Emeri’s editor. Read our editorial standards.

Get Emeri Weekly

A practical AI method, a worked example, and a task to try in your working week. One email a week.

We email you once to confirm. Unsubscribe in one click from any issue. See our privacy policy.

Related reading