Three passes · one meters per run · drafting is the ceiling

It works the inbox. You work the replies.

An inbox has two separate problems and everybody treats them as one. There is volume too much arriving, unsorted and there is obligation, the handful of messages that actually want something from you. An agent is genuinely excellent at the first. It cannot do the second, because deciding what you owe someone is judgement, and judgement is exactly what it cannot supply.

The three passes Volume vs obligation

3 passes, ~1 task each Volume is its job 3 things you keep Never the send button

This is the agent working your inbox. For composing and rewriting a single message, that is Meta AI and it is available far more widely.

Illustrative numbers. The shape holds: most of an inbox owes nobody anything.

If you only read one bit

Two tools, two halves of the same job

The short version

Muse does traffic control. Meta AI does composition.

Meta AI helps you write one message drafting it, tightening it, changing the register. Muse works the inbox as a whole: triaging what arrived, filing it by a rule you state, pulling details out of a thread into a note or a calendar entry. Different jobs, and most people who have both end up using both in the same ten minutes. What neither should do is press send.

Muse: sort, file, extract Meta AI: write, tighten, retone Drafts stay drafts Never delegate the send

Muse needs the Mail area granted before any of this runs, and an ungranted area produces a refusal that looks like a fault. How grants work · what the area actually covers.

The distinction the workflow is built on

Volume is a sorting problem. Obligation is not.

Think about what actually makes an inbox oppressive. It is rarely the writing most replies take ninety seconds once you have decided what to say. It is the not knowing: forty-seven unread, somewhere in there three things that matter, and no way to find out which without reading all forty-seven.

That is a sorting problem, and sorting is exactly what an agent is for. Pattern-matching across many short messages is also where the current model is strongest its tool-use scores went up in the last release even as its long-document accuracy went down.

The exhausting part of email is triage, not composition. Hand over the triage and the composition stops feeling like a burden, because you are only writing the three that mattered.

The second problem is different in kind. Deciding what you owe someone whether this needs a reply at all, whether it needs a warm one or a brief one, whether the honest answer is no is judgement about a relationship. An agent that offers to do it is offering to guess, and it will guess plausibly, which is worse than guessing badly.

So the workflow below is built around a single division: the agent produces lists; you make decisions. Every pass hands you an artefact to act on rather than an action already taken. That is not timidity it is the arrangement that actually saves time, because the reviewing is fast and the undoing is not.

One of these is a list. The other is a decision about a person.

The workflow

Three passes, and what to ask for each

One daily, one weekly, one when a thread demands it. Each costs roughly one task from the allowance, and each hands you something to read rather than something already done.

Daily~2 min

The triage pass

The one that earns its keep. You are not asking it to deal with anything only to tell you which handful of messages actually want something from you.

Go through my inbox from the last 24 hours and list every message that asks me a direct question I have not answered. Give me the sender, the question in one line, and nothing else. Do not draft replies, do not mark anything as read, do not archive anything.

What you get: A short list usually far shorter than the unread count, which is the point.
What you do: Decide each one. Reply now, reply later, or decide it does not need one. Three seconds each.

Weekly~3 min

The filing pass

Batched deliberately. The meter counts tasks rather than steps, so one weekly run over seven days costs a seventh of what running it daily would.

Label every message from the last seven days that contains an invoice, receipt or payment confirmation as Finance, and every message from a delivery or courier service as Deliveries. Do not move, archive or mark anything as read, and leave everything else alone. List what you labelled when you are done.

What you get: A filed week and a list of what moved, so you can spot a rule that is catching too much.
What you do: Check the list, not the inbox. If the rule misfired, adjust the wording before next week rather than re-running it now.

Ad hoc~1 min

The extraction pass

For when a single thread contains facts you need somewhere else. Narrow, checkable, and the closest thing to a free win in this whole workflow.

Read this thread and pull out every date, address and reference number into a new note called [subject]. Keep each one next to the message it came from. Do not add anything to my calendar and do not reply to the thread.

What you get: The details in one place, each traceable back to the message that stated it.
What you do: Glance and use it. Attribution is what makes it checkable in ten seconds rather than trusted blindly.

Careful with the very long thread

Summarising a fifty-message chain is the request people reach for first and the one this model handles least well. Independent testing of Muse Spark 1.3 found long-document accuracy slipped from 83% to 79% against the previous version, even as its tool-use scores climbed it got better at doing things and slightly worse at reading at length.

The practical consequence: many short reads across an inbox play to its strength, one enormous read does not. For a single long thread, Meta AI with the window attached is often the better tool.

Note the shape every request shares: a named scope, an explicit list of things not to do, and a finish condition. Drop the middle clause and labelling quietly becomes archiving. The four parts of a task.

Keep these

Three things not to hand over

Not because the agent would refuse because these are the parts where a plausible-looking mistake costs more than the time it saved.

The send

Meta says it confirms before sending, and you should want that even if it did not. Every other email mistake is recoverable; this one is read by a person before you know it happened.

Let it draft. You press send.

Deciding what a message deserves

Whether it needs a reply, whether it needs a warm one, whether the honest answer is no. That is a judgement about a relationship, and an agent will produce a confident-sounding answer without having the standing to make one.

It finds them. You weigh them.

Tone on anything that matters

Tone is the thing models get plausibly wrong rather than obviously wrong the reply reads fine and lands slightly cold, and nobody tells you. Read the draft as the recipient before it goes.

Fine on the page is not fine in the inbox.

Everything else the sorting, the labelling, the listing, the extracting is genuinely worth delegating, and it is most of the clock time an inbox takes.

FAQ

Email workflow questions

What is the difference between this and using Meta AI for email?

Meta AI helps you write one message drafting, rewriting, changing the tone. Muse works the inbox as a whole: triaging, filing, extracting across many messages. Composition versus traffic control, and most people want both. The full comparison.

Can Muse send emails for me?

Sending is the archetypal sensitive action and Meta says it confirms first. The workflow here deliberately stops at drafting, because an unconfirmed send is the one email mistake with no recovery. What to keep.

What is an agent actually good at in an inbox?

Volume sorting, labelling, listing, extracting, matching one pile against another. It is not good at obligation, which is deciding what you owe someone and in what tone, because that is judgement it cannot supply. The distinction.

Should I let it summarise a very long thread?

Be careful. Independent testing of Muse Spark 1.3 found long-document accuracy slipped from 83% to 79% even as tool-use scores improved, so a fifty-message thread is where it is weakest. Meta AI is often better for one long read.

Why does the daily pass ask for a list rather than drafts?

Because a list is reversible and twelve unwanted drafts are noise. The list is also the artefact you actually use you read it, decide, and most items turn out not to need a reply at all.

How many tasks does this workflow use?

Roughly one per pass, because the meter counts tasks rather than steps. Batching a week of filing into one well-scoped request is cheaper than running it daily, which is why the filing pass is weekly. How the meter works.

Will it mark my messages as read?

Only if you do not tell it otherwise, which is why every request here names what it must not do. Labelling, archiving and marking as read all feel adjacent to an agent, so state the boundary explicitly.

Is this safe for a work inbox?

That depends on your employer rather than on the app. A managed account may block third-party connections entirely, and routing work mail through a connection you authorised personally can cut across obligations your employer has. Managed accounts.

What if it drafts a reply in the wrong tone?

That is exactly why drafts stay drafts. Tone is the part a model gets plausibly wrong rather than obviously wrong, which makes it the worst possible thing to send unread.

Does this work outside the United States?

Muse launched in the United States only. Meta AI, which covers the composition half of this, has a far broader rollout and is available to most readers today. Where Muse works.

It produces lists. You make decisions.

Forty-seven unread, three that matter.

Finding the three is most of the work and all of the tedium. Answering them was never the part that made the inbox feel heavy.

The three passes Writing with Meta AI

Daily triage
Weekly filing
Ad-hoc extraction
Never the send

Independent guide not affiliated with Meta. Requests here are written by us as examples; statements about confirmation prompts are Meta’s own, and Muse is US-only at launch.