Two piles, then drafts

31 July 2026

Most of what your inbox costs you is not reading. It is deciding what each message deserves. You make that call forty times before lunch, and it is the reason an hour of email leaves you tired with nothing finished.

So separate the two jobs you have been doing at once. Sorting and answering are different work, and doing them together is what makes both slow.

First pass: sort only

Read my mail from the last two days and sort it into three piles: “needs me”
(only I can answer this), “for information” (read and move on), and “ignore”
(newsletters, noise). Give each message one line explaining why it is in that
pile. Do not draft any replies yet, only sort.

The last sentence is the one that matters. Without it the model starts being helpful and writes answers to mail you have not read. Now you are reviewing drafts for decisions you have not made.

The one-line justification is not decoration either. It is how you find out whether the sorting followed your logic or merely looks plausible, and the difference shows up precisely at the edges. An invoice that looks like a newsletter lands in the wrong pile silently. With a reason attached, you catch it in a glance.

Second pass: draft, for one pile only

Take the “needs me” pile and draft a reply to each in my voice. Short, direct,
no filler. Where you are missing information, put in square brackets what I
need to add. Send nothing.

“Send nothing” is there for the same reason as “do not draft yet.” The model drafts, I send. That division is what makes the whole thing usable: I read three drafts in two minutes and press send three times, instead of writing three emails from nothing.

What I had to undo

Two things, in the first week.

It moved a message from a client into “for information” because the body was a forwarded newsletter. The actual content was a line at the top asking about timing. The justification said “forwarded newsletter, no direct request,” which is how I caught it — the reasoning was visible and wrong.

And it drafted a reply declining something I had already accepted verbally. Nothing in the mailbox said so. That is not a model failure, it is a context failure, and it is the permanent shape of this problem: the agent sees your inbox, not your week.

The part I did not expect

After a week the justifications become their own report. They show you which kind of mail keeps landing in “needs me,” and it is almost always the same kind. Mine was scheduling.

That is a problem you can fix at the source rather than triage faster. The sorting was useful. Finding out what I was sorting was worth more.