Sixty-two overnight emails, sorted before you sit down

What is the exact prompt?

Paste this in the morning, adjusted to your own noise patterns:

• "Go through every email I received since 6pm yesterday. Sort each one into exactly one bucket: (1) needs a reply from me today, (2) needs a reply but can wait until Friday, (3) worth knowing, no reply needed, (4) noise."
• "For every bucket-1 email, draft a reply in my voice: direct, short sentences, no exclamation marks. If the email asks a question you can answer from the thread history, answer it in the draft. If it needs information you do not have, draft the reply with a [GAP] marker where I need to fill in."
• "Bucket rules: anything from my manager or a customer is never bucket 4. Anything mentioning money, contracts, or a deadline is bucket 1. Newsletters and automated notifications are bucket 4 unless they report a failure."
• "Output: one message listing each bucket with sender, subject, and a one-line gist. Leave all drafts as drafts. Send nothing, archive nothing, delete nothing."

What a good run looks like

A four-part list you read in ninety seconds, and a drafts folder already shaped. The [GAP] markers are the tell of an honest run: a companion that writes "I do not know which Thursday you meant" beats one that invents a date.

The common week-one miss is bucket inflation, where everything lands in reply-today. Fix it in the rules line rather than by re-explaining. "A vendor following up on an unanswered cold email is bucket 4, every time" kills the most common false positive.

How do you make it run every morning?

Save it as a routine once it survives a week of your corrections.

Same prompt, fixed time, triage list waiting when you sit down. Drafts still go through approval, and a scheduled run never widens what a companion may send. Ask Always in Settings → AI permissions tightens it further.

The deeper version, with rules that evolve and unsubscribe sweeps, is how to automate inbox triage. The tool surface is on the Gmail integration page, and Outlook users get the same shape through the Outlook integration.

Check these two things for two weeks

Open bucket 4 daily and scan the senders until it has earned trust.

A false noise call is the one error that costs you real money. Then read every draft before sending, because the voice takes a few corrections to land and correcting drafts is how it learns your patterns.

What you never need to check is whether something went out without you. Sends wait for approval whatever the prompt says, which is a different guarantee from a prompt-level request. The security model explains the distinction.

Frequently asked questions