How I Manage My Inbox (with AI)

I wrote about my inbox system a couple of months ago. The labels, the GTD discipline, the touch-it-once rule. Since then, the system has evolved.

The five labels are still the foundation: Needs Action, Awaiting Reply, Read Later, Remember This, and Delegated. Every email has a destination. Nothing sits in the inbox.

On top of that, I have filtering rules that run before I ever see anything. Some of these are standard stuff - unsubscribes, newsletter routing. But others are specific to how I run Rewind. Salesforce notifications about a deal that closed at $0 ARR go straight to trash. Google Docs comment notifications, trash. Stripe payout notifications, trash. The weekly sales summary from the sales team goes into Needs Action, because I want to read that one.

Writing those rules down forced me to articulate every category of email that arrives in my inbox and decide what should happen to it. I had been making those decisions implicitly for years. Getting them into the system made them explicit.

Enter AI agents

The bigger change is the triage agent. When I’m ready to process Needs Action, I run the agent and it works through the queue one email at a time. It reads the full thread, gives me a two-sentence summary of what’s being asked, and recommends an action: reply, create a Todoist task, delegate, move to Read Later, or archive. I approve or override. It executes.

The GTD rule still applies: every email gets an action. Nothing deferred to a pile you’ll never revisit. Touch it once.

What makes this work - instead of just being a slow way to process email - is that the agent knows my voice. I gave it a style guide covering my greeting format, how I sign off, my tone for different types of recipients, sentence length, what words I never use. It sounds excessive. But if the agent is drafting replies, they need to be replies I’d actually send.

Most drafts are close enough that I approve them as-is. The rest need a sentence changed. Either way, I’m not staring at a blank compose window trying to figure out what to say.

The part I underestimated was the upfront work. Not the technical setup - that was straightforward once the agent was running. The real work was naming every implicit decision I was already making. Why does this email get a reply and that one gets archived? What exactly is “Needs Action” versus “Read Later”? What is my actual voice - not in theory, but the specific things that make an email sound like me and not a template?

It took a few weeks to get it right. I’d notice something the agent got wrong, add a rule, and move on.

Now I don’t manage my inbox. I manage the exceptions.