How to get AI email drafts that sound like you

Start with examples you would send again

Choose your own writing that you are authorized to share with the assistant. For a manageable first pass, use four examples to describe your style and keep two other examples aside for checking. These counts are a starting scope, not a proven minimum. Remove signatures, contact details, quoted third-party messages and confidential facts that the task does not need. Label the audience and situation for each example.

For a new reply, supply the current thread separately, plus the points you want to make, facts you have confirmed and decisions you have not made. An old message can demonstrate phrasing; its delivery date, price or promise is not current context. A neutral greeting is a better fallback than inventing familiarity when the relationship is unclear.

Write a style card with an audience and exceptions

Ask the assistant to propose a short card you can edit. Use concrete rules you recognize rather than a personality label such as warm but professional. Require a reference to an approved example for each inferred rule, and mark guesses as proposals. Keep separate cards when you deliberately write differently to a teammate, a first-time customer or a friend.

Here is an illustrative card for ongoing project replies. It is invented for this guide, not extracted from a real inbox. The exception column keeps a preference from becoming a rule for every relationship.

Classify the edit before saving a preference

Suppose the current facts say Monday is a tentative target and approval is outstanding. An illustrative draft says: Dear Sam, I am delighted to confirm that I will definitely send the proposal on Friday. Warm regards. Your edited version says: Sam — thanks for the outline. I am aiming to send the proposal on Monday, pending approval.

That pair contains several kinds of correction. Ask the assistant to show the differences and propose what, if anything, should carry forward. You decide which rules to save. A polished draft with the wrong commitment still fails.

Check a new reply with and without the card

Use a thread you kept aside. Give the assistant the incoming request and the facts that were available at the time, but withhold your final reply. Request one draft without the card and another with it, using separate fresh conversations when possible. Supply the card only to the second conversation. Saved memory or global instructions may still carry over; if you cannot exclude that context, treat this as an informal comparison rather than a clean baseline. Keep the assistant, task facts and other instructions the same. Compare both with your own writing and judgment; do not ask the assistant to award itself a voice score.

Use one reserved teammate reply and one for a different audience, with an unresolved date in either. This is a small personal check, not a benchmark or proof of future reliability. If a reserved example is used to revise the card, choose a new example for the next check.

1. Check meaning first. Are the names, dates and amounts supported? Did it answer the actual questions? Did it add a promise, opinion or relationship detail you did not supply?
2. Check audience next. An ongoing teammate thread may allow a brief opening; a first customer introduction should not inherit that familiarity. Ask yourself whether you would actually send the wording.
3. Check the approved rules. Identify specific lines that fit or miss the card. Record needs a fact fix, needs an audience fix, needs a style edit, or ready for my review; keep the reason beside the draft.
4. Compare the two versions. If the card makes a reply worse, narrow or remove the responsible rule. Keep any improvement claim limited to the examples you inspected. Review the next few real drafts yourself.

Keep a feedback log you can undo

Keep each approved rule beside its example reference, audience, decision date and card version. Return the assistant’s draft and your final edited draft explicitly; changes you make in another app may be invisible to it. Ask for proposed updates, then approve, narrow or reject them. Do not silently rewrite the card after every edit.

For example, log E-04 → short greeting → teammate thread → approved in version 2. Log E-05 → never use a greeting → rejected as too broad. Record corrections to proposal dates in that reply’s context, not the style card. These are illustrative records, not observed results.

Save the current card and retain the prior version. When a rule causes trouble, mark it retired and explain why. At the next session, ask the assistant to identify the card name, version and audience it can actually access. If storage is unavailable, paste the current card and examples into the conversation. No durable memory is needed to try that manual version.

Keep the finish at a draft you can inspect

Start with one reply returned in chat or a document. If your assistant has a supported, authorized email connection, ask it to save a draft in the chosen thread and verify the draft is there. Google’s Gmail documentation distinguishes creating a draft from sending it. Neither a saved draft nor a good style match authorizes sending.

When revising a saved draft, check its current contents first and preserve your edits. Keep its thread and draft reference so a retry updates the intended draft instead of creating another. If the assistant cannot read or save drafts, return the text for you to paste manually. Review recipients, attachments, facts and tone before you send.

For Muse, Instinct, Dots or Grok Bot, use the capabilities available in your own account. The first trial needs only text input and drafting. Stored cards, email access and recurring triage are separate capabilities to verify. This guide does not claim that any of these assistants automatically captures edits or shares the same memory behavior.

A starter brief: Draft one reply to [thread] using [current facts] and [approved audience card, version]. Keep examples separate from current facts. Flag anything missing. Return the reply and any proposed card changes separately. Do not send. After I supply my edited version, classify the changes and wait for my approval before saving a lasting preference. The linked inbox prompt below supplies a fuller workflow when you are ready to add triage.

What the sources establish—and what they do not

In a Grok Bot Builder Demos video published October 1, 2026, Shub Gaur describes editing YapBot’s email drafts and varying his voice by audience. We rechecked the public captions on October 3. This is a vendor-affiliated practitioner’s report, not our own execution test or evidence of a particular learning mechanism.

Jason Vana’s June 17, 2026 article describes comparing drafts with edited versions before approving changes to a client voice guide. His writeup also promotes his services. We credit that feedback approach without adopting his quality promises. The worked examples, audience exceptions, reserved-example comparison and reversible card process here are our untested editorial adaptation.

A September 22, 2026 preprint by Shintaro Sakai and colleagues reports that AI drafts shifted participants’ style in a controlled workplace-email study. Its authors note that participants did not send real emails and that lasting effects remain unknown. That is a reason to inspect whether a draft preserves your intent, not to infer anyone’s preferred style from their nationality.