Create an executive inbox copilot

Step 1 — Grant reach, read-only on purpose

Connect email. What just happened, mechanically: the connector renders your inbox as text the window can read, and your requests as actions the tool understands. The reference entry’s rule now becomes operational: a connector that reads is a research assistant; one that sends is an actor. This lab never lets the AI send anything. First, calibrate what it can see:

Read my inbox from the last three days. Don’t summarize yet — just tell me: how many messages, how many are from actual humans, how many are automated, and the date range you can actually see.

Step 2 — Design the triage; don’t accept the default

Generic triage is the average inbox’s triage. Yours gets designed:

Interview me about how I actually handle email: what kinds of messages I get, which I dread, what slips through, what truly matters. Then propose a triage system of 4–5 categories defined by WHAT THE MESSAGE NEEDS FROM ME — for example DECIDE (only I can make this call), ANSWER (needs a reply), WAITING (I’m owed something), READ LATER, IGNORE. Refine it with me until the definitions are sharp enough that two different people would file the same message the same way.

“What it needs from me” is the executive move. An inbox isn’t a pile of text; it’s a queue of demands on your attention, and most of them are counterfeit.

Step 3 — Run the triage

Triage the last three days against our system. For each message: the category and a one-line why. For every DECIDE item, state the decision I’m actually being asked to make in one sentence, with its deadline if one exists. Return a table, DECIDE first.

This is the whole lab in one beat: the copilot’s job is not to answer email. It’s to surface the three real decisions buried under eighty messages — and to state them so plainly you can’t defer them by re-reading.

Step 4 — Drafts for review, never sends

For every ANSWER item: draft the reply in my voice. Here are two emails I’ve written that sound like me: [paste them]. Put all drafts in front of me. Do not send anything.

Two real examples steer voice better than any list of adjectives. Then the ritual: read each draft, edit, and send it yourself. The gate isn’t friction — the gate is the design. It’s the difference between a copilot and an intern with your signature.

Step 5 — Make it an agent, honestly

Look at what you just ran: read → classify → draft → wait for the human. That loop is what an agent is — this one with you as the scheduler. If your tool can schedule recurring tasks, schedule the daily triage. If it can’t, save the whole thing as a project or reusable prompt called “morning triage” and fire it with one message over coffee — same loop, human clock.

Either way, ask the agents entry’s four questions of your own creation, out loud: What’s in its window? What can it touch? Where does its truth live? Who reviews its work before it counts? And keep the boundary permanent: the copilot never gets send. The day you’re tempted, re-read its draft to your most sensitive thread and ask whether you’d stake that relationship on an unreviewed average.

---

Check your understanding: why does the copilot get read but never send — and what is its actual job, if not answering email?

Use this process for other incoming work. Read → classify by what-it-needs-from-me → draft-for-review → human gate. That’s the Slack firehose, the support queue, the contract pile — every inbox-shaped stream in your life just became triageable.