Set up an AI email assistant that writes like you

The inbox routine

Twice a day at most. Morning, and late afternoon if the inbox is heavy.

• Ask for the sort. Ask Claude which threads need a reply from you today, which need your thinking, and which are only information.
• Ask for drafts. For the reply group, ask for a draft of each reply using your voice skill.
• Edit and approve. Read each draft, fix what is off, then send it from Gmail or approve it in Claude if your connection allows sending.
• Fix the file. When a draft gets you wrong in the same way twice, add a rule to the voice file.

Can Claude send from Gmail? Anthropic's pages disagree

Anthropic's help center says Claude can search and read your email, draft replies, and send, reply and forward, and that by default it asks for your approval before each of those actions (help center, checked Oct 21, 2026).

Anthropic's Gmail page in the Claude docs says the connector is read-only and cannot create, send or modify messages (Claude docs, checked the same day).

The routine above works either way. If your connection cannot send, ask for the draft in the chat and paste it into Gmail.

The voice file, line by line

Make one document and fill in each line, the way you would brief a new assistant on their first morning.

• How I open. For example: first name and a comma, no Hope you are well.
• How I close. For example: Thanks, then my first name. No title, no full signature in replies.
• Normal length. For example: three sentences or fewer for scheduling and quick answers.
• Words I never use. Your own list. For example: leverage, circle back, per my last email.
• Who gets the formal version. Name the clients or the situations, such as a first email to a lawyer or a banker.
• What never goes in writing. Fees for a new client before a call, anything about another client, anything you would not want forwarded.
• Examples. Paste a starting set of your sent emails, unedited, for example 20 to 30, across the kinds you write most: a scheduling note, a quick yes, a polite no, a follow-up after a meeting, a reply to a worried client.

Turn the voice file into a skill

Save the voice file inside a skill so Claude uses it without being asked. Claude loads a skill when your request matches the skill's description, so describe it around replies and drafts. The full steps, and a ready description for a reply-in-my-voice skill, are in Claude skills, explained with one complete example.

Skills need code execution turned on. On Pro and Max that is under Settings > Capabilities.

Outlook and Microsoft 365

Anthropic's pricing page lists Claude for Microsoft 365 and Claude for Microsoft Outlook on its paid plans (Claude pricing). This page does not cover what that connection can do, such as sending or attachments. Check Anthropic's Microsoft 365 documentation before you rely on it. The voice file and the routine stay the same.

Where it gets things wrong

• Attachments. Through the Gmail connector, Claude sees attachment metadata, and the content inside the file stays out of reach (help center). If the answer is in a PDF, add the file to the chat.
• Context outside email. It does not know about this morning's phone call. A draft can be confident and wrong. Read every one.
• Client data on consumer plans. Anthropic says it does not train its models on Gmail, Drive or Calendar connector data. On Free, Pro and Max, if you allowed your chats to be used for training, content you copy from those connectors, or Claude's responses that include connector content, may be used to improve its models (help center). Check that setting before you put client email through Claude.

Questions