Draft and rewrite emails that still sound like you

How to use a plain chat window to draft, rewrite, and tighten emails across every situation you face, without losing the voice your relationships are built on.

What you'll have when you're done

A repeatable approach for every email situation that eats your time: the tricky decline, the sensitive renewal conversation, the cold intro, the overlong draft you wrote in a hurry. You paste, you tell the model what matters, you get a draft in your voice in under a minute. You copy it, tweak two words, hit send.

No Gmail connector. No terminal. Plain chat window, today.

The AI email that sounded like a press release

You know the one. You asked Claude or ChatGPT to draft a reply, you were in a rush, you sent it, and the recipient responded with "wait, did you have your assistant write this?" or just replied with noticeably shorter sentences than usual because the vibe shifted.

The draft was technically correct. It said the right thing. But it opened with "I hope this message finds you well," used passive voice in three places, and had a closing line that sounded like it belonged in a vendor contract. The relationship took a small, invisible hit.

The problem is not the model. The problem is that you never told it how you actually write. Without that signal, it defaults to a generic professional register that feels like a press release precisely because it is optimized for nobody in particular.

The fix is five minutes of setup and one paste.

What you need first

• A Claude or ChatGPT account. Plain chat, free tier works for this.
• Two or three emails you actually sent. Typical ones from the last month: a reply to a customer, a note to a vendor, a check-in with a hire. Not your best work, not the one you agonized over. The ones that just sounded like you on a normal day.
• Nothing else. No integrations, no setup, no configuration.

If you want to get to this faster, the first 30 minutes with Claude is the right primer. But you can also just open a chat window and start.

Step by step

The voice-matching setup move (do this first, every time)

This is the unlock. Before you ask for any draft, paste two or three emails you actually wrote and say:

"Here is how I write. Match this voice for emails in this conversation."

That single instruction beats any tone adjective. "Direct but warm" means something different to the model than it does to you. Your actual emails carry your sentence rhythm, your typical sign-off, whether you open with context or cut straight to the ask, whether you use contractions, how long your paragraphs run. Paste three of them and the model reads all of that at once.

This is the same show-don't-tell logic from using examples in prompts. An adjective is guesswork. An example is a pattern. The pattern is what the model actually runs on.

Without the setup, every draft you get will drift toward formal-generic. With it, the first draft usually needs only two or three word changes before it is ready.

Do this at the start of each new conversation. The model does not carry it forward across sessions automatically, though you can make it stick by saving it to a Claude Project, which holds standing instructions and examples so you do not have to re-paste them each time.

The five email situations

1. Reply to a thread

Paste the thread. Then say what you actually want to convey, not what you want to write. Something like: "I want to say yes to the Q3 timeline but push back on the Q2 ask. Do not apologize for the delay. Keep it under 150 words."

The specifics about what to convey matter more than any tone instruction. Tell it the substance and the model figures out the framing.

2. Decline while keeping the relationship

Paste the ask. Then give the model what it cannot infer: who sent it and what the relationship is worth. "This is a two-year agency partner. We are not moving forward with this project but I want to keep the relationship warm. Give me a warm no that leaves the door open. No hedging, no fake 'we'll revisit in Q4' language."

The model will not know the relationship history unless you tell it. Give it the context it needs, or the draft will be technically fine and contextually wrong. Giving AI the right context goes deeper on this if it clicks.

3. A tricky or sensitive message

Describe the situation plainly, not the email you want. The cleaner the situation description, the better the draft.

For example: "Two-year customer pushed back on renewal price. They are a strong reference account. I want to hold the price but I absolutely do not want to lose them. Give me a draft that acknowledges their concern, holds firm on price, and pivots to the value conversation."

The model writes prose, not decisions. You own what the email says. The draft is a starting point; for anything sensitive, read it out loud before you send it.

4. Cold or warm intro

Give it the four things it needs: who the recipient is, what they do, what the connection is, and what you want. Then add a length constraint. "Three sentences, direct, no flattery opener, no 'I came across your work and was impressed.'"

The opener is where AI email goes wrong most often. It wants to open with a compliment. You probably do not. Say so.

5. Tighten a draft you already wrote

Paste your draft. "Tighten to under 120 words. Keep my voice. Do not add apologies or softeners I did not put there."

This is the highest-value use of the whole workflow. You wrote the thing. The model just cuts it to the bone without losing the substance. What used to take you two editing passes takes ten seconds.

Tone controls worth having

These work as add-ons to any of the five jobs above. Paste them at the end of your instruction when you need them:

• "Warm but firm, not apologetic."
• "No exclamation points."
• "Under 120 words."
• "Match the directness of my samples."
• "No 'hope you're well' opener."
• "No passive voice."

You do not need to explain these. They are precise enough that the model follows them reliably.

The three-versions move

If you are unsure of the tone or you want to compare options, end any request with: "Give me three versions: one direct, one warmer, one shorter."

You will pick one or combine pieces. This is faster than iterating a single draft back and forth. Three options in ten seconds beats five revision cycles in five minutes.

How you'll know it's working

The first draft needs minimal editing. Not zero edits, but two or three word changes rather than a full rewrite. If you are rewriting more than 30% of the draft, the setup is the problem: the model does not have enough of your writing to work from, or you described what you want rather than what you want to convey.

People stop noticing. The best sign is that nobody asks if you had help. The email sounds like you. The relationship holds the temperature it was at before you hit send.

You stop dreading the hard ones. The tricky decline, the sensitive price conversation, the cold intro to a board candidate you do not know: these used to sit in your drafts folder for two days. They stop sitting there.

When it breaks

The draft is right in content but wrong in context. This means you gave the model the what but not the who. It does not know your history with the recipient, what you promised them last quarter, or that they just lost a key employee and are unusually stressed. You have to tell it those things. Paste the context the same way you paste the emails: briefly and plainly. "Important context: we had a rough call three weeks ago about delivery timelines and this is the follow-up."

Names, numbers, and dates are off. The model does not have access to your calendar, your CRM, or the thread above the one you pasted. Proofread every name, date, and commitment in the draft before you send. This is not optional. A draft that says "your April delivery" when the date was March is worse than a late reply.

For something genuinely sensitive, the draft is a starting point, not the answer. A message about someone's employment status, a real conflict with a partner, anything with legal weight: draft it here, yes, but then read it out loud. Check every sentence for what it actually commits you to. The model is good at producing something that sounds right; you are the only one who knows whether it is right. You own the final words.

The model sounds too cautious or over-hedges. It is protecting you from perceived risk. Override it directly: "Drop all hedging language. This is a clear yes, say it like one."

It is not improving across sessions. You are rebuilding the voice setup every time. Save your sample emails and the setup instruction to a Claude Project so the model has your voice loaded at the start of every conversation.

Level up

This workflow lives entirely in chat. The next step is connecting the model directly to your inbox so it can read threads, write drafts, and sort what deserves your attention. That is a more involved setup, but the ROI is larger: instead of pasting threads one by one, you run one instruction and get a sorted inbox with drafts already waiting. Inbox zero for CEOs walks through that workflow from the ground up.