You will be able to turn your own rough notes into a first draft that keeps your content and judgement.
When you ask an AI assistant to write an email and give it almost nothing to go on, it fills the gaps with guesses. The draft can read smoothly and still say things you never meant to say. The fix is to split the work: you decide what the email says, and the assistant decides how to say it.
Before you open the assistant, write your points as rough bullets: short notes that hold the substance of the message. They don't need to be full sentences or in any order. They do need the facts, your position and, most importantly, the decision or request you are making. Once those exist, only the wording and the order are missing, and wording and order are what the assistant is good at.
Take Jun Wei, an example whose details are invented. On Thursday afternoon she has a reply she has been putting off. A client's HR manager wants to move a two-day workshop forward by a week. The trainer is not free that week, and the only alternative dates fall in the client's busy month. She types into the assistant: "write a reply to this client about rescheduling". The draft is polite and well organised, but wrong. It offers a date she cannot offer, apologises for something that was not her company's fault, and suggests a discount nobody approved. The assistant had almost nothing to work with, so it guessed.
She starts again and writes bullets in her own shorthand:
can't do the week they asked, trainer away two options: keep the original dates, or first week of the following month no change to the fee either way need their answer by Friday next week to hold the trainer offer a quick call if easier
That is about forty words, written in two minutes. It covers what she can and cannot do, the options, the deadline and the next step. She pastes the bullets in, with the voice guide from lesson 3.1 already saved in her project, and asks for a reply. Saving a voice guide in a project makes drafts come back in your own style. Her draft comes back in her style, uses both of her options and nothing else, and needs one small change to the opening line.
The split between you and the assistant is a division of labour. The assistant handles wording, order and tone. You control content: what is said, what is offered and what is promised. This matters most for anything that commits you or your company, such as prices, discounts, deadlines, delivery dates, and what is or is not included.
When those commitments come from your bullets, the draft can only rearrange them. When you leave them out, the assistant often fills the space with something that sounds reasonable, and reasonable-sounding commitments are what get people into trouble. Read every draft for anything that was not in your bullets, and delete every new promise.
Bullets also carry your judgement. You might know that a client values speed over polish, that a contact is new and nervous, or that the last email went badly. A short note such as "she's new, keep it reassuring" passes on context the assistant cannot get any other way.
The method for any message runs like this. Before opening the assistant, write rough bullets with the facts, your position and the decision or request. Add any commitments yourself. Add judgement notes about the recipient or the situation where they help. Paste the bullets and ask for a draft, ideally with a saved voice guide. Then read the draft for anything not in your bullets and delete every new promise.
For a reply, paste the incoming email together with your bullets. The assistant then knows what it is replying to and can answer each question in the order the sender raised it.
Remove personal details from the incoming email first, using the method from lesson 2.4. Names, phone numbers and signatures can almost always go, because most replies need the content of the message rather than the identity of the sender. Keep the email and your bullets clearly separated. Simple labels work: "Their email" above one block, and "My reply should say" above the other.
Then check that your bullets answer every question in the incoming email. If the email asks more questions than your bullets answer, the assistant may answer the extra ones for you. Say an email asks four questions and your bullets answer three: the assistant may write an answer to the fourth, and that answer will be its guess.
For anything longer than an email, such as a proposal, report or policy document, one big request tends to produce a long, even, slightly vague document. Every section gets the same weight, and the important parts get no more attention than the boilerplate. Lesson 3.1 of Working with AI assistants, "Why one giant prompt gives you a mediocre result", explains why the approach below works better.
Long documents work better in two stages: agree an outline first, then draft section by section. Give the assistant your notes and ask for an outline, meaning a list of section headings with one line on what each covers. Edit the outline until it matches what you want. Then draft one section at a time, giving the assistant that section's bullets plus the agreed outline for context, and read and fix each section as it comes back. Checking is easier this way, because you work through one page at a time while you still remember what that section was meant to say.
Pick three emails you need to send this week. Write the bullets for all three before you draft any of them. That is a good first test of the method.
Take three emails you need to send this week, write bullet points for each, and draft them with your voice guide applied.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).