I rebuilt my email in Claude

In April I built an AI operating system over a long weekend, and it still runs my days. Email was one piece of it, and that piece worked. Every morning it read across my inboxes, sorted everything, and drafted the replies.

But it only happened once a day, and the drafts landed back in each account's drafts folder, so I still had to go to four different places to send them. Anything that arrived after breakfast I handled by hand. Most of what the system gave me was a list of things to deal with later.

Last month I rebuilt it. Twice a day now, everything anyone has asked me for arrives on one screen, across four email accounts and a Teams channel, already sorted into things that need a reply, things that need me to do something, and things I only need to know. The replies are written and the folder is chosen. I read each one, change whatever I want to change, and approve it. Within the hour it has gone out from the right account with my signature, and the original is filed.

The part I got wrong the first time was the hand-off. When I gave Claude a whole piece of work rather than an email, it disappeared, and I had to remember to go and chase it, which defeated the point. If I had accepted that version I would have stopped using this by now. Now it becomes a real task on my task list, owned by the agent, and I can watch it move from queued to running to done.

I still edit almost everything it writes. That is not a complaint about the system, it is the system: it keeps every draft it suggested, compares it to what I actually sent, and writes down what it learns. Most of what it has learned is not about grammar. It is that my drafts kept handing decisions back to other people when the decision was already mine to make.