Sort your inbox with Claude: draft, never send

Every email goes into one of four labels:

• REPLY. It needs your answer, so Claude writes a draft.
• TO DO. There is a task inside, so Claude pulls it out with the deadline.
• READ. Updates and newsletters get one line each.
• SKIP. Promos and cold pitches are listed and left alone.

Why four and not ten? Because each label matches one thing you do next: answer, act, read or ignore. More labels look tidy, but they give Claude more chances to guess, and they give you more sorting to check. Four is enough to clear a morning inbox.

Most worries about AI and email come down to one fear: a message goes out that you never saw. So Claude may write, but only you send. Enforce that rule twice.

1. 1

   In the prompt. The prompt below says draft only, and says never send, delete, unsubscribe or archive.

2. 2

   In the tool. Claude's Gmail connection can send, reply and forward, and by default it asks you first. One tired click on approve and a half-right email is gone. So switch the sending tools off too. The Gmail connector lists three that send: send_message, reply and forward.

The order matters. The tool settings only appear once Gmail is connected. So:

1. 1

   Add the Gmail connector.

2. 2

   Straight away, go to Customize, Connectors, Gmail, Tool permissions, and set send_message, reply and forward to Blocked.

3. 3

   Only then run the prompt.

As of October 2026, connector settings can change, so check Anthropic's connectors page if your screen looks different.

Examples make the drafts better. Paste three emails you have already sent. Claude copies your length and your tone.

Pick three different kinds: a quick reply, a longer explanation and a polite no.

Choose emails that sound like you on a normal day, not your most formal ones. Claude copies what you give it. If all three samples are stiff, every draft will be stiff. Remove names, phone numbers and anything private from the samples before you paste them.

Paste emails in, or let Claude read them if your email is connected.

Replace the two curly-brace lines at the bottom of the prompt with your real text. Keep each sample email separate, so Claude can tell where one ends and the next begins. If you paste a long batch, start with fifteen or so. A short batch is easier for you to check line by line.

Picture a small design studio with six people. The owner opens her inbox on Monday and finds twelve new emails. She pastes her three sample emails, then pastes the twelve. Here is a short slice of what could come back:

REPLY
1. From a client asking if the logo files are ready.
Draft: Hi, thanks for checking in on the logo files. ASK ME: what delivery date can I give them?

2. From a client unhappy about a late proof.
I'll handle this one.

TO DO
- From the print shop: send them the revised proof (no deadline in the email)

READ
- Software vendor: new version notes, nothing needs changing

SKIP
- A marketing agency: "Grow your studio this quarter"

Look at what happened. The first draft sounds like her, and it did not invent a delivery date. It stopped and asked. The second email is a complaint, so Claude labelled it REPLY and left it for her. That is the third rule in the prompt doing its job.

She answers the ASK ME question in the chat, tells Claude to adjust the draft, then copies the final text into her email and sends it herself. The task line goes on her own to-do list. The SKIP list takes ten seconds to scan. The whole pass takes her a few minutes instead of a long start to the day.

Run it first on made-up emails, or a handful of low-risk ones. Check that:

• Sensitive emails say "I'll handle this one" and have no draft
• Drafts that lack a fact say ASK ME
• Nothing was sent
• Every email shows up once. Count the emails you gave it, then count the items across all four labels. If the numbers do not match, ask Claude to list every email it read, with sender and subject, and find the one that went missing.

A sensitive email that also asks you to do something still lands under REPLY with "I'll handle this one". Rule 3 wins, so its task will not show up in TO DO. Add that task to your own list when you handle the email.

Run the same test a second time with the same emails. The labels should land in much the same places. If they jump around, your label descriptions are too loose, so tighten them.

Also check the TO DO lines against the original emails. Claude can misread a date, so confirm each deadline yourself.

Common mistakes

1. 1

   Running it before blocking the send tools. Block first, run second.

2. 2

   Pasting three formal samples. The drafts come back stiff. Use emails that sound like a normal Tuesday.

3. 3

   Starting with your real inbox. Test on made-up or low-risk emails first. You will see where the labels go wrong before it matters.

4. 4

   Trusting the SKIP list without a glance. A real client email can look like a pitch. Scan the sender and subject lines once.

5. 5

   Never changing the prompt. If Claude keeps labelling meeting requests as READ, add a line: meeting requests always go to REPLY. The prompt gets better each time you correct it.

Your morning now starts with drafts, not a full inbox. Read each one before it goes. You are still the sender.

When you read a draft, look at three things: any fact you did not give Claude, any promise, and the tone toward that person. Claude writes what sounds likely, and it can be wrong. A draft that reads well is not the same as a draft that is right.