Claude Code as a Chief of Staff: 39 Runs, and Where Claude Duet Wins
You can make Claude Code your chief of staff, and we did. Ours is a command each founder runs at the end of the day. It reads the day’s meetings and email, files what matters into the right client folder, finds replies we owe, proposes tasks and lines up tomorrow. Between June 4 and September 30 it ran 39 times across my co-founder and me, read 460 meetings and 1,420 email threads, and surfaced 626 threads where someone outside the company was waiting on us. The short version of how to build one is at the end: a folder, one command and two weeks of running it by hand.
The weak spot is in that first number. Thirty-nine runs over the 85 working days from June 4 to September 30, split between two people, means each of us ran it on about one working day in four or five. A chief of staff who only shows up when you remember to call them misses the days you were too busy to remember.
That gap is what Claude Duet is built to close. Anthropic opened a waitlist for it on October 21, for Pro and Max accounts in the US. It checks your accounts on its own and sends you a morning brief. I haven’t used it yet, so what follows is our system in detail, Duet as Anthropic describes it, and an honest call on where each one wins.
What our chief of staff does in one run
The command is /strategy:day-end-process, the most-edited of the commands our company runs on, about 1,100 lines of instructions plus helper agents and scripts. The main session never reads a transcript or a full email itself. It pulls headers, sorts them by client, and hands the raw text to small agents that each work in their own context and return a short structured result. That keeps the run from drowning in a day’s worth of email. (How Claude gets at our Gmail in the first place is in Can Claude read my email?)
At the end it prints a summary. Here’s the shape of one, with the client details taken out:
DAY-END PROCESS COMPLETE
MEETINGS + EMAILS: 6 meetings, 19 threads reviewed, 36 tasks created
NEEDS YOUR REPLY: 2 threads owed, 2 reply drafts created
TASK REVIEW: 1 merged, 1 marked done, 3 overdue pushed
DOC UPDATES: 7 structured edits applied
CONTACT / CRM: 22 contacts synced, 4 outreach drafts, 9 clients flagged at risk
TOMORROW'S MEETINGS + PREP: [time, meeting, what to prepare]
TOMORROW'S FOCUS: [top three things to actually do]
Those numbers come from one real run in June. Yes, 36 tasks from one day, which comes up again below. Over the four months the runs also made 303 doc edits, synced 270 contacts and researched 57 companies.
Why ours runs at night instead of in the morning
A morning brief is the obvious design, and Duet has one. We went the other way.
The day is complete. At 9pm today’s meetings have transcripts and today’s email has replies. A morning run is mostly summarizing yesterday, and anything that changed overnight is noise until someone acts on it.
It sets up tomorrow. The last phase pulls tomorrow’s calendar, flags meetings that need prep (a client call with no recent notes, an open proposal, a first intro), and picks the top three things to do using a simple filter: eliminate, automate, delegate, then do what’s left. Waking up to that list is better than waking up to a list of what happened.
It needs a person in the chair. Nothing gets created until the founder approves it. The run shows one table of proposed tasks and possible duplicates, and waits for a reply like “create all,” “drop 4” or “merge 2+5.” Contact updates have their own approval step. That’s a conversation, and a conversation works better at a desk at the end of the day than on a phone at 7am.
What it never does is send. Every reply it writes lands in Gmail as a draft, and a person reads it and presses send (why we hold that line).
The numbers, including the embarrassing ones
Show the data
| Month (2026) | Runs | Mine | My co-founder’s |
|---|---|---|---|
| June (from the 4th) | 8 | 3 | 5 |
| July | 10 | 5 | 5 |
| August | 14 | 7 | 7 |
| September | 7 | 3 | 4 |
We built for the missed days. Every run writes a line to a ledger, and the next run reads it. If the last run was more than a day ago, it widens the sweep to cover the gap, up to five days back. Two runs did exactly that. It’s a patch on the real problem, which is that the system only works when someone starts it.
The other number I’m not proud of is 1,042. That’s how many tasks the runs created in those four months, and most of them were noise. Telling the model to bundle related tasks didn’t fix it. A script that does the grouping did (the full story). Then there was the same client thread showing up in both founders’ inboxes, which both runs wanted to file. That needed a shared table and a rule about who owes the reply (two founders, one brain).
What Claude Duet is, and who can get it
Claude Duet is an Anthropic app that works for one person in the background, checking their accounts on its own schedule. It’s a research preview for Pro and Max accounts in the US, through a waitlist, with invitations going out in waves. It isn’t on Team or Enterprise plans yet. During the preview it costs nothing extra, and it uses your plan’s usage limits, including the five-hour and weekly limits.
What it does, from Anthropic’s getting-started guide and its safety article, not from my own use:
- Each morning it reviews your email and calendar, picks out priorities, and can start on first steps like drafting replies for your review.
- It checks your connected accounts at set intervals through the day, so something urgent can still land between checks.
- It works on its own cloud computer with its own browser, and returns bigger jobs as files like decks, financial models and dashboards.
- It asks before it sends, pays, signs or deletes anything. Trading securities and moving money between accounts are blocked even if you approve.
- It runs on Opus 5.5 with no model choice.
Where Duet wins
It shows up without being asked. That’s the whole problem our ledger is patching. A chief of staff that checks in on its own will catch the day you forgot.
There’s nothing to build. Ours is months of work in a git repository, plus someone who can read a Python error. Duet is a waitlist and an app.
It lives on your phone and can run errands. It has a browser and one-time cards for purchases. Ours has neither, and isn’t trying to.
Where building wins
Two people can share one brain. Duet belongs to one Pro or Max account. Ours reads both founders’ mailboxes into one repository, knows when my co-founder already filed a thread, and still keeps track of who owes the reply.
The rules are files everyone can read. “AI never sends client email” is a ruling in a log, with a date and a history, loaded into every run on every laptop (write down the rulings). In a personal agent, the equivalent is a preference in one person’s account.
What it learns stays with the company. Every filed thread, synthesis and task lives in storage the company owns. When someone leaves, the knowledge stays. Duet’s memory is tied to the person’s account, and it keeps what it learned until that person resets it.
Approval is enforced in code. Our drafter has no code path that sends. Duet asks first by design, and Anthropic is clear about which approvals you can’t turn off. Both are reasonable. Only one of them is something you can read line by line.
How to make Claude Code your chief of staff
Start much smaller than ours. You need a folder, one command and a habit.
- A folder with a
CLAUDE.mdthat says where things go: one subfolder per client, atasks.md, adecisions.md. - One command in
.claude/commands/day-end.md, roughly:
Read today's calendar and the email threads where I'm waiting
on someone or someone is waiting on me.
1. For each client touched today, append three lines to
clients/<name>/log.md: what happened, what we promised, what's next.
2. List every thread where someone outside is waiting on me.
Draft replies as Gmail drafts. Never send.
3. Propose tasks as a table. Create nothing until I approve.
4. Look at tomorrow's calendar. Tell me what needs prep,
and the three things I should do first.
- Run it yourself for two weeks at the end of each day, before you automate anything. You’ll find out fast which steps you actually read and which ones are noise, and that’s cheaper to learn by hand. Ours created 1,042 tasks before we learned it.
If after two weeks you’re forgetting to run it, that’s real information. Either schedule it (what Anthropic’s scheduled tasks can and can’t do) or get on the Duet waitlist and let it come to you.
Which one I’d pick
- Just you, mostly your own email and calendar: try Duet when your invite arrives. You’ll get most of what a chief of staff does with none of the setup, and you can always build later.
- A team that needs the same answers: build. A chief of staff for a company has to read several people’s accounts, remember what the company decided, and survive someone leaving. That’s what we built, and it’s what the build sets up for teams in 14 days.
- Both: fine, as long as company accounts get connected through company tools. What I’d check before connecting your books to any agent is its own post.
To see what a chief of staff with your company’s context feels like, try the demo. It’s a made-up agency with a Zoom call from this morning. Ask it to draft the proposal for the person on that call and read what comes back. If you’re a founder weighing this for your own company, here’s how I’d set it up for you.
