You don't sort a full inbox by reading through it. You sort it by assigning every message exactly once to one of five piles: Now, Today, Can wait, Trash, and Judgment. The assignment isn't made by feel but by three questions in a fixed order: does anyone have to do anything at all? Is someone waiting on it today? Can I answer it without asking back? Each pile gets its own treatment and its own slot in the day, otherwise sorting is just a rename of the problem. The fifth pile matters most: anything your rules don't clearly cover goes there and nowhere else. Write that down once and you no longer have a sense of order, you have a procedure someone else can run.
I built my office so that sorting and replying are separate passes, because the expensive part isn't the typing, it's the same decision made over and over at the start. How I take a workflow apart before changing anything is on the Process optimization page. Which tasks are suitable for handing off and which stay with you is covered in Handing tasks off to AI.
Sorting isn't tidying, it's deciding
The effort in a full inbox isn't in reading and isn't in replying, it's in looking at the same message three or four times and starting the decision from scratch each time. Microsoft names exactly this in its Outlook guidance and sets the goal of reducing the number of times you touch each message (support.microsoft.com). The same page gives the Four Ds (Delete, Do, Delegate, Defer) and the two-minute rule: if it takes less than two minutes, do it now, otherwise defer it.
From that follows what most folder guides skip: a folder decides nothing, a folder is a place. If you drag a message there without having decided when and how it gets handled, you've only moved it out of sight. That's why these are piles: a pile is a treatment class with a fixed processing time. Folder, label or category is only the technical representation of it.
The five piles
The last column is missing from the guides and decides in practice whether the system holds. A pile becomes useless the moment it turns into a catch-all, and "Now" only helps as long as it stays rare.
One distinction that regularly gets lost: trash is not archive. Only what you have vetted and what can be retrieved at any time gets deleted automatically, for instance a routine notice that stays readable inside the tool. System notices are the trickiest case, because they also carry security alerts, failed payments, delivery failures and cancellations. Anything of business relevance, including the copy of a finished matter, is therefore not deleted but archived under a written retention rule. Which retention periods apply to you is something you settle with your tax adviser or your lawyer; this post describes workflow organization and is not legal or tax advice.
Three cases missing from many grids run inside the piles here. Delegating is Today: passing it on is the answer you can give without asking back; after that it becomes a case of waiting. Waiting on others is Can wait, but only with a date on which you follow up; without a date, waiting isn't a pile, it's a hiding place. Archiving without action is question one answered no, and still not trash.
The three questions, in this order
The order is part of the procedure, not a matter of taste: the cheapest decision comes first.
- Does anyone have to do anything at all? No means trash or archive, with no further thought. In most mailboxes this question clears the largest group, in seconds. Skip it and you end up rating advertising by urgency.
- Is someone waiting on it today? Not how urgent the message sounds, but whether something actually goes wrong today if nobody acts: deadline, appointment, blocked handover, payment. Yes means Now, no moves you to question three.
- Can I answer it without asking back? Yes means Today. No means Can wait, if it's clear what's missing. And Judgment, if not even that is clear.
What the three questions achieve is unspectacular and effective for exactly that reason: they replace a verdict ("is this important?") with a check ("is there a deadline today?"). A verdict can't be written down and can't be handed over. A check can.
The six rules that carry the system
- One pile per message. No multiple assignment, even when two would fit. The moment a message sits in two piles, it effectively sits in none.
- Touch once, decide once. What that means is: no repeated reading without a result, not a frozen decision. If you disagree with yourself on second look, usually the message isn't the problem, the rule is, so change the rule. And if new information arrives, a deadline moves or you filed it wrong, you reassign it as a matter of course.
- Every rule names an observable feature. Sender, subject pattern, attachment type, a date in the text. Not: "if it's important." Someone else can check a feature, nobody can check importance.
- The doubt pile is mandatory. Without it, unclear cases drift upward, because uncertainty feels like urgency. The fifth pile keeps that pressure out.
- Every pile has a time, not just a name. Now is now, Today is a block, Can wait is a weekly slot, Trash is never. A pile without a processing time is a warehouse.
- Sorting and replying are two passes. Start answering while sorting and you sort the rest worse, because your head is already inside the first message.
How to write your rules down
The real step isn't the sorting, it's the writing down. A rule has three parts: recognizable feature, target pile, reason. The reason isn't decoration, it's what later tells you whether the rule still holds. The rows look like this:
Two things matter more than the rows themselves. First: you recognize a deadline from the subject or the text, never from the sender. A tax adviser also sends same-day queries and passes on official deadlines, which is why the row above names the known monthly report and not "sender plus PDF". Second: a feature is a hook, not proof. The question mark misses the indirect request ("please get back to me by Friday") and flags the rhetorical question along with it. Where a feature doesn't hold, the three questions decide, and in case of doubt the judgment pile.
Five to fifteen such rows are my starting value, not a measurement: in my mailbox they cover most of it. The judgment pile catches the rest, and every second or third case there turns into a new row. That's the whole learning process: you don't write everything down up front, you write down what has come up more than once. The first rows are the hardest. If you'd rather not sit over them alone, put your list up in the community, in a call or as a post, and have it read by people who have written their own.
One note on the technical side: if your rule has a hard feature, your mail client can run it for you. Gmail officially recommends labels and filters rather than rigid folders, because one message can carry several labels (Gmail Help). That only appears to contradict rule 1: several labels serve finding things again, the pile stays one.
Where the built-in tools end and a role of its own is needed is settled by four criteria in AI Email Triage: Agent or Tool?, along with what Outlook, Gmail, Apple Mail and Thunderbird can do by themselves.
What I take from Inbox Zero and the four-folder method
Both are older and better known than anything here, and both are inside the system. Just not completely.
When someone else does the sorting
This is where writing it down pays off, as with each of the five workflows in office organization for the self-employed: a list of feature, target pile and reason is a work instruction. You can hand it to a person or to an AI employee. In both cases the same cut applies: the sorting goes away, the answering stays with you. Skip that cut and you haven't handed off the sorting, you've handed off your voice.
That's how Frieda works in my office. She sorts the inbox into exactly these four piles plus the judgment pile, puts drafts in place for the Today pile, and never sends anything herself. Private mail, job applications and health matters she doesn't touch, not even to sort them. That isn't a promise, it's a matter of setup: those messages are separated out before anything reads them at all. There are four common ways to do it: a separate mailbox for the private part, a folder outside the connected selection, a sender list, or a server-side rule that files such mail away on arrival. What the connection doesn't see, it can't read. Receipts are a way station for her, not an archive. The interesting part isn't that an AI can do this, it's that it doesn't work without these rules and without that separation.
Which of the four ways is easiest for you depends on your provider; you set it up once. A good question for the community, in a call or as a post: there are people there who have the same connection behind them.
What those rules look like word for word is in CLAUDE.md: Claude Code's Memory: Frieda's complete house rules with task, rules, boundaries, files and approvals. And how a sorted inbox feeds into the next step, from receipt to quote, is shown in AI Agent for Bookkeeping and Inbox.
Frequently asked questions
Should I sort my emails by date or by sender?
Both are views of the list, not systems. By date helps when working through a backlog, by sender helps when pulling one matter together. Neither takes the per-message decision off your hands: set the view so you don't have to jump around while sorting, then decide by the three questions.
Do I need a separate folder for every pile?
No. A pile is a treatment class, not a storage location. Folders, labels, flags or categories all work, as long as one pile has exactly one counterpart. Just don't create more destinations than you have piles.
What do I do with the three thousand old messages already sitting there?
You don't run those through this system, but you don't sweep them away wholesale either. Back the stock up first. Then go through it in a separate, risk-based pass: known deadline senders first, then unread messages, authorities, contracts, payments and security alerts. Anything still open goes into the normal piles. Only the remainder gets archived in bulk and stays findable through search. That's an appointment of its own, not something done in passing; in the community there are people who have that pass behind them.
How often should I sort the inbox?
Twice a day is my starting value, not a measurement: in my office that's enough, and once is enough more often than expected. What matters anyway isn't the frequency but that sorting is its own short pass. If you look in between, you're looking, not sorting.
How do I know my pile system isn't working?
Two signs. First: the Now pile fills up every day, which means the rules are worded too softly and let gut feel through. Second: the judgment pile stays empty. That almost never means everything is clear, it means unclear cases are being filed somewhere else, usually upward.
Where to go from here
Don't start with the inbox, start with a blank sheet: write down five rows, feature, pile, reason, taken from the messages you already handled this week. Then sort for a week using only those five rows plus the three questions, and add a row every time the same doubtful case shows up a second time.
Once that list exists, the question is no longer whether you can hand this off, but to whom. Frieda is available as a ready-made package built exactly this way in my community Claude Practitioners, together with the other AI employees and the course.
