You will be able to get targeted edits on your own drafts for clarity, length and tone.
When you ask an AI assistant to "improve" something you wrote, it rewrites it in its own voice. The result is often smoother, but it stops sounding like you, and it can drop or change things that mattered. You can also use an assistant as an editor. Used that way, it makes the specific changes you ask for and shows you what it wants to change before it changes anything.
Jun Wei writes client proposals. One evening she wrote a proposal cover note herself, because it mattered and she wanted it to sound like her. It ran to 340 words (an invented example, like the rest of her figures), and she knew it was too long. She pasted it into an assistant and asked it to "improve" the note. The new version was shorter and smoother. It also lost the sentence about the client's previous workshop, which was the reason the client would say yes. And it added "we are committed to delivering exceptional value", a phrase she would never write.
So she went back and asked for a list of cuts. The assistant gave eight suggestions, and she accepted five. She rejected the one that removed the sentence about the client's previous workshop. She also rejected two that softened her deadline. The note came out at 230 words and still sounded like her.
A good human editor does not read a draft once and fix everything. They read it several times, with one question each time. Is it too long? Is the request clear? Will the reader misunderstand anything? You can ask an assistant to work the same way. Each request, or pass, names one job. For a proposal cover note, three passes might be:
Cut the draft by a third without losing any facts or the request. Make the request in the last paragraph clearer, so the reader knows exactly what to do and by when. Read the draft as the client's finance manager and flag anything they might misread or object to.
The third pass is often the most useful, because it finds problems you cannot see in your own writing.
Run one pass at a time. If you ask for shorter, clearer and friendlier all at once, you get a different document, and you cannot tell which change came from which request. Lesson 4.2 of Working with AI assistants, "Feedback that gets a better second draft", made the same point about revising AI drafts. It applies just as much to your own writing.
The most useful change you can make in how you ask is this. Request a list of suggested changes with a reason for each, and don't accept a rewritten version. A request might read: "Do not rewrite this. List up to eight changes that would cut this draft by a third. For each one, quote the original words, give the suggested replacement and say briefly why."
Then go down the list and accept or reject each item. The draft stays yours, and only the changes you agree with go in. You get two benefits from working this way. First, you control every word, so the voice stays yours. Second, you learn. After a few weeks of reading suggestions with reasons, you start to see your own habits. Common ones are a throat-clearing first sentence that delays the point, a request buried in paragraph three, or the same word used four times.
Some suggestions look like improvements but are not. These are flattening edits. They make the writing less like you or less effective, and they come in three common types.
The first type is added filler. Phrases such as "I wanted to reach out", "please feel free to" and "we are committed to" make the writing longer and less like you.
The second type is a softened request. "Please confirm by Friday" becomes "it would be great if you could let us know at your convenience". The softer version is politer, but it is much less likely to get an answer by Friday. Jun Wei rejected two suggestions like this, because they softened her deadline.
The third type is a change in meaning, and it is the most serious. Say "we can usually deliver in two weeks" is edited to "we deliver in two weeks". The email now makes a promise you did not make. Or say "subject to trainer availability" is dropped. You have then lost a condition that protects you. Before you accept any suggestion that touches a fact, a number, a date or a commitment, compare it with your original.
Whatever you accept, read the complete final version before sending, from the first word to the signature. When suggestions are applied one at a time, a sentence may no longer fit the one before it. A sentence may also refer to something that was cut.
You can't skip this final read, because the email goes out under your name. If it promises something, misstates a figure or reads oddly to the client, you are accountable for it. "The assistant edited it" is not an answer anyone will accept.
To practise, choose one of your own drafts, ideally one of the three from lesson 3.2. Run it through two editing passes using the methods above. Give each pass one job, ask for a list of suggestions with reasons, accept or reject each one, then read the whole thing through. Keep track of which suggestions you turn down and why.
Run one of your own drafts through two editing passes, accept or reject each suggestion, and note which kinds of edit you rejected.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).