Reverse Neutralization: How to Make an LLM Write Like You
Every email a language model drafts for me sounds like it was written by a committee that has never met me. Correct, polite, and completely interchangeable. For a long time I assumed that was the price of delegating. Then I found the fix in Generative AI Design Patterns, by Valliappa Lakshmanan and Hannes Hapke, filed under a name I liked before I understood it: reverse neutralization. The trick is odd enough that I wanted to write it down.
Neutralization, briefly
Let’s start with the boring direction, because you need it to understand the useful one.
Neutralizing a text means removing the voice and keeping only the information. Same facts, zero person behind them. Here’s a before-and-after:
Original
The new coffee machine is SO good, my mornings are basically fixed now, thanks HR you’re the best ☕
Neutralized
The new coffee machine is a nice upgrade and has improved my mornings. Thanks to HR for arranging it.
Every fact survives: there’s a new machine, it’s good, mornings are better, HR gets the credit. What died is the person. The caps, the slightly dramatic “my mornings are basically fixed”, the cup. Nobody reacts to the second one.
And here’s the uncomfortable part: the second one is exactly what models produce by default. Not because they can’t write, but because “polite and inoffensive” is the safest average of every email they’ve ever seen. You didn’t ask for the neutral version. You got it for free.
Reverse neutralization
Now read the example backwards. Start at the polite nothing, end at the message a colleague would actually send.
That’s the direction I care about. I don’t need a model that takes my emails and flattens them; my own brain does that at 6 pm on a Friday. I want one that takes a correct but lifeless draft and gives it a pulse. Not necessarily the caps and the emoji, but the sense that a specific human wrote it.
The obvious move is to just ask. “Make it warmer.” “Less formal.” “More human.” And it doesn’t work, or it works in the worst possible way, because the model has no idea what you mean by human. Its idea of warm is whatever the internet calls warm, which in practice means it adds “I hope this finds you well” and an exclamation mark. Your idea of human is specific. It’s how you open, how you handle bad news, whether the ask goes in the first line or the last. You can’t describe that in an instruction. You probably can’t describe it at all. You can only point at it.
So point at it.
The dataset trick
To get a model to make the trip from neutral to voiced, you’d ideally show it examples of exactly that trip: here’s a bland email, here’s the same email with a person in it. Pairs. Input and output.
The problem is that those pairs don’t exist in nature. Nobody writes the corporate version of their email, then the real one, and keeps both. You will never find that dataset.
But you can manufacture it. Backwards.
Original
The new coffee machine is SO good, my mornings are basically fixed now, thanks HR you’re the best ☕
Neutralized
The new coffee machine is a nice upgrade and has improved my mornings. Thanks to HR for arranging it.
Step 1: Collect the originals. Find ten emails that sound like you: ones you wrote on a good day, or ones you draft for the occasion.
Spread them across situations: a follow-up, an apology, a no, a welcome, some bad news. Voice shows up differently in a thank-you and in a payment reminder. Copy-pasting by hand is fine and arguably better than sending an agent to hunt for them, because you’ll be pickier.
Step 2: Neutralize them. Hand the emails to a model and ask it to rewrite each one as a standard, impersonal business email, keeping every fact and losing everything else. This is the one task models are naturally excellent at, so let them shine. You want these to be really bad and really complete: anything the neutral version drops, the model will later learn to invent. Here’s what that looks like on one of mine:
Original
Subject: Welcome aboard, Tom
Hi Tom,
You start with us on Monday and I wanted to write before then, so you don’t walk in completely blind.
A few practical things: your desk is the one by the window (you got it because Paul moved floors, not on merit: that will come). Your laptop is already set up, I’ll give you the password in person. At 9:30 we’re having coffee with the team, no slides, no formal introductions.
Don’t expect to understand everything in the first week. Nobody has, me included. Ask questions, even the ones that feel silly: those are usually the best ones.
See you Monday,
Fabio
Neutralized
Subject: Welcome to the team
Dear Tom,
Ahead of your start date on Monday, I wanted to share some practical information so that your first day runs smoothly.
Your workstation is located by the window, which became available following Paul’s move to another floor. Your laptop has already been configured; login credentials will be provided to you in person. An informal team coffee is scheduled for 9:30 am, with no presentations or formal introductions.
Please note that a full understanding of our processes is not expected during the first week. Should you have any questions during your onboarding, however minor, do not hesitate to ask.
We look forward to welcoming you on Monday.
Best regards,
Fabio
Every fact survived, even Paul: the desk, the laptop, the coffee with no slides, the permission to ask silly questions. What died is me. The dig about merit, the “nobody has, me included”, the feeling that someone who remembers their own first day is on the other end. Which was the only reason to send the email at all. Perfect. That’s the input we want.
Step 3: Flip the pairs. Take each neutral version and use it as the input, and the original as the output. No training involved: the pairs go into the prompt as examples, and the model makes the journey in the direction you need, even though you built the data in the opposite one.
That’s the whole trick. You taught the model to write like a person by first asking it to write like a template.
The prompt
Here’s the version I use, with three of my ten pairs already in it. Copy it, swap in your own, done.
You turn neutral emails into emails with a specific voice.
Keep the information intact; change only how it's said.
Don't use emoji, don't add facts that aren't there.
Examples:
INPUT:
Subject: Update on the outstanding report
Dear George,
I am writing to inform you that the report originally scheduled for delivery yesterday has been delayed. The data cleaning process required more time than initially anticipated.
The complete report will be delivered tomorrow by 1:00 pm. Should you require it earlier, the August sales section has been finalized and can be provided upon request.
I understand you have a meeting scheduled and apologize for any inconvenience this may cause.
Kind regards,
Fabio
OUTPUT:
Subject: I'm late, and here's how I'll make it right
Hi George,
I'll say it straight away: the report you were supposed to have yesterday isn't ready. I don't have a good excuse. I underestimated how long cleaning the data would take and realized it too late.
Here's what I'll do: you'll have the full report tomorrow by 1 pm. If the August sales section would help in the meantime, that part is done and I can send it right now.
Sorry for the trouble, I know you had a meeting coming up.
Fabio
INPUT:
Subject: Request for a brief conversation
Dear Clara,
I hope this email finds you well. I am currently looking into procurement processes within hospital organizations, and I understand you spent approximately two years working in that sector.
Would you be available for a 20-minute conversation within the next two weeks? A phone call would be sufficient, at a time convenient for you. I am aware that this is a busy period for you.
Please let me know whether this would be possible.
Thank you in advance.
Best regards,
Fabio
OUTPUT:
Subject: A favour (a small one, I promise)
Hi Clara,
I'm writing because I need a hand and you're the right person to ask.
I'm trying to understand how purchasing works inside hospital organizations, and I remember you spent a couple of years in that sector. Would you have twenty minutes in the next two weeks for a chat? Phone is fine, whenever suits you: I know you're in a busy stretch.
If you can't, no problem at all, just tell me and I won't push.
Thanks either way,
Fabio
INPUT:
Subject: Reminder: outstanding invoice no. 47
Dear Mr Marini,
We are writing with reference to invoice no. 47, dated 15 July, which remains outstanding. The payment due date was 30 August.
Should there be any discrepancy relating to the amount or to the invoice itself, please let us know so that the matter can be resolved.
A copy of the invoice is attached for your convenience.
Thank you for your attention to this matter.
Kind regards,
Fabio Cingottini
OUTPUT:
Subject: Invoice no. 47: a gentle reminder
Dear Mr Marini,
A quick note about invoice no. 47, the one from 15 July. It was due on 30 August and I haven't seen the payment come in yet.
My guess is that it simply slipped through, it happens, especially in August. If instead something on the invoice doesn't look right to you, the amount or anything else, tell me and we'll sort it out together. I'd rather fix it than chase you.
I've attached a copy so you don't have to dig for it.
Thank you, and have a good day,
Fabio Cingottini
Now transform:
INPUT:
[the neutral draft]A few notes on using it.
Three examples are enough to make the model change gears, and five is noticeably better. Past that you’re mostly spending tokens for small returns. That’s why you collect ten and use a handful: pick the pairs closest to the email at hand, a no for a no, an apology for an apology, and leave the rest in the drawer.
The “no new facts” line matters more here than anywhere else. When a model is asked for personality it tends to either decorate or invent, and an invented deadline or a meeting that doesn’t exist, in an email that goes out under your name, is not a style problem. It’s a liability. Block it and the voice has to come from structure, rhythm and word choice, which is where it should come from anyway.
Notice that the instructions are deliberately boring. All the taste lives in the examples. If you want a different voice tomorrow, drier, warmer, more formal for a new client, you swap the pairs and leave the instructions alone. The one exception is the emoji line: if your originals use them, drop it, or the examples and the instructions will argue.
One caveat: the output is exactly as good as the originals you picked, so if your ten emails are mediocre you’ve built a very efficient mediocrity machine.
Otherwise, go make some boring emails. On purpose, just this once.