How to Use Claude to Write Emails (Without Sounding Like a Robot Wrote Them)

Most people use Claude for email the same way, and it is the way that produces the worst result. They open a blank chat, type "write a professional email declining this meeting," and paste whatever comes back.

What comes back is fine. It is also generic, slightly too long, and does not sound like you, so you spend four minutes editing it and conclude the whole thing was not worth the trouble.

The trouble was not Claude. It was that you asked for an email instead of giving it the two things it needed to write yours: the raw content, and a sample of how you actually write.

Here is the setup that fixes it, plus the prompts to lift.

Start with the reply, not the request

The single highest-leverage change: stop asking Claude to compose from a description, and start giving it the actual thread plus your rough answer.

You already know what you want to say. You just do not want to spend eleven minutes shaping it into something sendable. So say it badly, fast, and let Claude do the shaping.

Prompt to lift:

Here is an email I received:

[paste the full email]

Here is my rough answer, unedited:

[type your answer in whatever shape it comes out. Fragments are fine.
Say the real thing. "no, too expensive, but ask me again in Q1" is enough.]

Rewrite my answer as a sendable email. Keep my meaning exactly, including
anything I declined or committed to. Do not add warmth I did not write.
Do not add offers I did not make. Match the length of the email I received.

That last constraint does more work than it looks like. A three line question deserves a three line answer, and left alone most models will return nine.

Teach it your voice once, use it forever

The reason Claude-written email sounds like Claude is that you have never shown it what you sound like. One paste fixes that, and if you set it up as a Project, you never paste again.

Prompt to lift:

Below are five emails I wrote myself. Read them and describe my writing
style back to me: sentence length, how I open, how I close, how formal I am,
words I use often, words I never use.

[paste 5 of your own real sent emails. Pick ordinary ones,
not your best ones. Ordinary is what you actually sound like.]

Read what it gives you back. If the description is wrong, say so and give it two more samples. If the description is right, save it, because that paragraph is now a reusable instruction.

Then attach it to everything:

Write in this style:

[paste the style description it produced]

Never use em dashes. Never open with a question. Do not use the words
[list any word that makes you wince].

Put those instructions into a Claude Project and every email you draft in that Project starts from your voice instead of the default one. We covered the Project setup separately in 7 Claude Projects for small business.

The four email jobs worth automating

Not every email benefits from this. These four do, because each has a shape you repeat constantly.

The decline. The one people put off for days. Give Claude the request, your reason, and whether the door stays open.

Decline this request. My reason: [one line]. The door is [open / closed].
Three sentences. Warm but not apologetic. Do not over explain the reason.

The follow up on silence. The email nobody enjoys writing because every version sounds needy.

I sent the email below [X] days ago and got no reply. Write a follow up
that is two sentences, adds one new piece of value or context, and does
not guilt them or apologize for following up.

[paste original]

The long thread summary. Fifteen replies, four people, and someone new needs to be brought in.

Summarize this thread for someone joining now. Give me: what was decided,
what is still open, and who owes what. Under 150 words. Do not editorialize.

[paste thread]

The difficult one. The email you have rewritten six times and still have not sent.

Here is what I want to say, unfiltered:

[write it exactly as you feel it, including the parts you would never send]

Rewrite this so it is direct and professional, keeps every substantive point
including the uncomfortable ones, and removes only the heat. Do not soften
the actual position.

That last prompt is the one people report back on most. It works because you were never actually stuck on wording. You were stuck on tone, and tone is the thing a model is genuinely good at separating from content.

The check before you hit send

One rule, applied every time, and it takes fifteen seconds.

Read the email and ask whether every fact in it is one you supplied. Models fill gaps smoothly. If you did not give it a date, a price, or a name, and one appears in the draft, it was invented to make the sentence flow.

This is the only real risk in the whole workflow, and it is fully solved by looking. Facts you supplied, keep. Facts that appeared on their own, cut or verify.

What this actually saves

Say you write twenty emails a day that require thought, and this takes each one from four minutes to ninety seconds. Run that math on your own count rather than trusting mine, because the number that matters is yours.

The bigger change is usually not the minutes. It is the four emails sitting in your drafts folder that you have been avoiding since Thursday. Those get sent now, because the hard part was never the typing.

Want the prompts without copying them out of a blog post?

Every prompt above, plus the ones that did not fit, are in the free prompt pack. No cohort, no call, no drip sequence.

Get the free prompt pack

FAQ

Do I need a paid Claude plan for this? No. Everything above works on the free tier. The Projects setup, where you save your voice instructions once and reuse them, is the piece a paid plan makes smoother.

Will people be able to tell? Not if you seed it with your own rough answer and your own voice samples, which is the entire point of the first two sections. Generic in, generic out.

Is it safe to paste a client email in? Treat it exactly like any other cloud tool you paste client information into. If your business has a rule about that, the rule applies here too. If it does not have a rule, that is worth a conversation before it is worth a prompt.

Can I do this in Outlook or Gmail directly? There are ways to wire it in, and they are worth it once the prompts above are earning their keep. Start in the chat window. Wire it up after you know which four emails you actually repeat.

What if the output is still not right? Give it one more real sample of your writing rather than another instruction. Samples beat adjectives every time.