Lesson 2 of 8 · 8 min read · last verified 2026-08-26
Email without losing your voice
In this lesson you will:
- Draft routine replies without adopting a generic assistant tone
- Recognise which emails should never be drafted for you
Email is the obvious first candidate. It is frequent, it is formulaic, and you can usually tell at a glance whether a reply is right.
It is also where AI-assisted writing is most recognisable, and where being recognised costs you something. So this lesson is as much about restraint as about drafting.
Ask for your voice, not a professional one
The single most common mistake is asking for a “professional” or “polished” email. Both words push the model towards the average of all business correspondence it has ever seen — which is why the output arrives padded with I hope this email finds you well, I wanted to reach out, and Please don’t hesitate.
Nobody writes like that. Everybody recognises it.
The fix is the one E3·L3 taught: supply context instead of adjectives. Paste two emails you actually sent, and ask for a reply that matches them.
Below are two emails I’ve sent. Match their length, directness and level of formality. No opening pleasantry — I don’t use them.
Reply to the message below: agree to the Thursday meeting, ask them to send the figures beforehand.
The difference is immediate, because you have replaced a vague quality word with evidence.
Say what the email must do
A reply has a job. Say what it is, in the imperative, before anything else: decline without closing the door; confirm and add one condition; ask for the thing they forgot.
Then name what must not happen. Do not apologise. Do not commit to a date. Do not repeat their message back to them. Negative constraints work well here because email has strong conventions the model will otherwise follow by default.
Two sentences you write yourself
Some emails should never be drafted for you. Not because it would not work, but because the effort is the message:
- Bad news. Redundancy, rejection, a project cancelled. The recipient will read it several times.
- An apology. An apology you did not compose is not one.
- Disagreement with a person, as opposed to a position.
- Praise. Specific, personal praise is worth more than fluent praise, and the difference is audible.
A useful test: if the recipient would feel worse on learning it was drafted, write it yourself. That test is doing real work, because they may well find out — and increasingly, they will guess.
Check before you send, every time
A drafted email is written in your name. If it invents a date, misspells a person’s name, or agrees to something you did not intend, none of that is the assistant’s problem — it lands as your word.
So read it as the recipient, not as the author. Every name, every number, every commitment. This takes twenty seconds and it is not optional.
Try it now (6 minutes)
Find a routine email you sent recently. Paste two of your own sent messages, then ask for a reply to it in your voice with the pleasantries removed.
Compare it with what you actually sent. Note the specific words the assistant reached for that you never would — those are the ones to name and forbid next time.
Check your understanding
Recap
Give the assistant your own emails rather than an adjective, state what the reply must do and must not do, and never delegate bad news, apologies, disagreement or praise. Then read every draft as the recipient before it goes out in your name.
🗂 3 flashcards from this lesson join your daily review.
Previous: Where AI actually fits in your day · Next: Reading long things faster