Email that sounds like you
- prompting
An AI-drafted email fails in exactly two ways: it does not sound like you, and it promises things you never agreed to. Both are fixable with the same structure: show your voice with examples, and bound the facts with placeholders.
Voice is examples, not adjectives
“Professional but friendly” describes a thousand different writers. Two emails you actually sent describe one. Paste them as voice samples — the same move as few-shot examples, applied to yourself: openers, how you phrase requests, how you sign off. The draft inherits your cadence instead of a generic one.
The four-part request
- Outcome — what the email must achieve: apologize, explain the fix, confirm next steps.
- Voice samples — one or two real emails you have sent to similar people.
- Facts and boundaries — what must be in it (the refund, the correct date) and what must not be promised (no discounts, no future credits).
- Placeholders — “where a detail is missing, insert [refund reference] instead of inventing it”.
That last rule is the quiet hero. A model asked to “write the email” will happily produce a reference number, an amount and a deadline that you might send by accident. A model asked for placeholders gives you a draft with visible gaps — and gaps you can see are gaps you can fill.
A bad example
Ask: “Write a professional apology email to a client about a billing mistake.”
Generic tone, invented specifics, and a promise about “adjusting charges” you are not authorized to make.
A better example
Ask: “Write a reply about charging the client for a fourth storage week after their goods shipped — our mistake. It must apologize plainly, state that the week will be refunded, and give a next step. Match the voice of these two emails I sent before: [paste]. Do not promise anything beyond the refund. Where a detail is missing, insert a bracket placeholder like [refund reference].”
Same speed, your tone, and no commitments you did not make.
Before you send
Read it once as the recipient — a rubric pass: does it answer their question, does it promise anything you cannot deliver, is every bracketed placeholder filled in? That is the five-question check applied to correspondence — thirty seconds, no exceptions for “just an email”.
Keep the voice pack current: every few months, replace the samples with more recent emails. Voice drifts — yours included — and two fresh samples cost nothing while keeping the drafts fitted to the person you are now.
Practice
Assemble a voice-matching email prompt
A client was charged for a fourth storage week after their goods had already shipped — a billing mistake on your side. You need a reply that apologizes, explains the fix, and does not promise compensation you are not authorized to give.
Fill every field to assemble a prompt for a reply email that sounds like you — then compare with a strong example.
A strong example (self-assessment — not machine-graded)
Write a reply email to a client who was billed for a fourth storage week after their goods had already shipped — our mistake. The email must: apologize plainly, state that the fourth week will be refunded, and give a clear next step with a date. Match the voice of these two emails I have sent before: [paste example 1] [paste example 2] Do not promise anything beyond the refund — no discounts, no future credits, no service-level commitments. Where a detail is not in my notes, insert a bracket placeholder like [refund reference] for me to fill in before sending. Context: [paste the facts of the case]
Hint
Adjectives like 'friendly' cannot be checked; examples can.
What convention keeps the model from inventing an amount, a date or a reference?
Why this is the answer
Two examples replace a page of tone adjectives, and the boundaries line fences off the promises you cannot make. The placeholder rule leaves visible gaps — [refund reference] — instead of plausible inventions you might send by accident. A draft you must complete is exactly the right shape for consequential email.
Transfer
- Recurring replies: keep a two-email voice pack ready to paste.
- Sensitive topics: more boundaries, more placeholders, no drafting shortcuts around a human review.
- Any language: samples work wherever you write.
Next
Next: Extract data from documents — the same discipline, pointed at piles of text that should become rows.