Teaching AI to Write Like You - Yuval Peles

What’s in this guide?

Twelve steps for training AI to write in your voice. The process, the mistakes along the way, and a final checklist for every writing session.

Most people I meet open a chat, type “write me a post about X,” and get something that could have been written by any business coach in the world. They read it, they don’t connect to it, they go back to writing by themselves. Or worse, they publish it and wonder why nobody reacted.

It doesn’t work that way for me. And it took a long time to understand why. The AI doesn’t know you. It doesn’t know who you are, how you talk, what you never say. When you ask it to write, it goes to the average. And by definition, the average doesn’t sound like anyone.

In this guide I’ll walk you through the process I went through, with all the failures along the way, until the AI actually started writing in my voice. It’s not a thirty-minute process. But once you’ve done it, every post turns from an hour’s work into ten minutes.

One thing worth saying up front: you’re going to fail along the way. Those failures are exactly where the system gets built. The most meaningful moment in this whole process for me was the day I asked for five posts and the AI produced five bad ones. That’s the moment that turned everything. I’ll get to that.

Chapters in this guide

Why “write me a post” doesn’t help
Collect the material you already have
“Read it, and tell me what you understood”
The first post will come out weak. That’s the most important stage.
The moment that changed everything
Force it to read examples, not just rules
The “don’t” list is worth as much as the “do” list
Correct, save, repeat
Two agents, not one
The structure that grows on its own
When the AI says “yes” without actually doing it
The final checklist for every writing session

Why “write me a post” doesn’t help

People open the chat and type “write me a Facebook post about why it’s important to stop procrastinating on decisions.” They get back a 200-word response that starts with “We all know procrastination is one of the greatest challenges…” and ends with “Remember, small steps lead to big changes!”

You read it, and there’s an uncomfortable feeling that’s hard to name. “That’s not me.”

Why isn’t it you? Because the AI doesn’t know who you are. It defaults to what it does know. The average of everyone who has ever written a post about procrastination. And the average of every writer isn’t a person. It’s a ghost.

So what do people do? They improve the prompt. They write “write me a Facebook post about procrastination, in a personal voice, without marketing language.” They get the same post, slightly different.

The principle: The problem isn’t the prompt. The problem is that the AI doesn’t know you. Another line of instructions won’t fix that. What will fix it is infrastructure. Real material. A process that takes time.

Collect the material you already have

Before you ask the AI for anything, there’s one piece of work to do. Work everyone wants to skip. Not skipping it is the difference between a generic output and one that’s actually yours.

Take everything you’ve ever written. Everything. Posts you published on social, transcripts of talks you gave, articles you wrote for a blog, emails. Even long WhatsApp messages that felt like real writing. Put them in one folder.

When I started, I collected 29 posts I’d written over the years. I dumped everything into a folder. Then I went to the AI and said one very clear thing:

“Read all of these. Tell me what you understood about how I write. Don’t write a post yet.”

The last line is the critical one. “Don’t write a post yet.” Because the AI wants to write. It’s an output-generating machine. If you let it jump to output, it will jump. And then you have a post based on a shallow read.

What to do: Open a folder. Put in at least 10 examples of writing in your voice. If you have 30, great. If you only have 5, start with 5 and add more later.

“Read it, and tell me what you understood”

The AI comes back with an analysis. Now people tend to skim it, say “yeah that sounds like me,” and jump straight to asking for a post. Don’t do that.

Take the analysis and go through it slowly. Read sentence by sentence, and ask yourself: “Is that really me? Or is that going in a different direction?”

You’ll find three kinds of things:

Yes, that’s me. The AI caught you right. Check it off.
No, that’s not me. The AI thought you were something you’re not. Write down why not.
That’s missing. Things it didn’t catch, that you know are you.

Go back to the AI with corrections. “Yes, yes, no. I never use that phrase. And something’s missing: I tend to end paragraphs with a rhetorical question. Update your analysis based on that.”

Why this matters: This analysis is the AI’s memory of you. It’s going to be saved, pulled up, and used as the basis for every future post. If the analysis is 10% off, every post will be 10% off. Until you fix the analysis.

The first post will come out weak. That’s the most important stage.

Now you ask for a post. One. Not five. One.

Give the AI a topic, maybe a bit of context, and ask for one post in your voice. It’ll write. And I can almost guarantee you: the first post will come out weak.

That’s not a failure. That’s the stage. And more than that: if the first post comes out great, something is off. Because the AI hasn’t gotten enough corrections from you to actually write like you. It read an analysis (even a correct one). Reading an analysis isn’t practice.

Instead of “didn’t love it,” phrase specific observations:

“The second sentence starts with ‘It’s important to note.’ I never use that.”
“The third paragraph is 4 lines, then another 4, then another 4. My rhythm shifts.”
“It wrote ‘a message worth internalizing.’ I would never write that. I’d write ‘something that sits with me.'”

Every observation like that is valuable material. It becomes a correction. It goes into memory.

The principle: You’re not building a post. You’re building memory. Every correction is a brick in the wall. After 5 to 10 corrections, you’ll have a writer. After 20, an excellent writer.

The moment that changed everything

I’ll tell you about my moment. Because it’ll happen to you too, and it helps to know what to look for.

I’d been at it for two months. The AI’s memory was loaded. Most posts felt fine. I was sure the system was standing.

One day, the evening before a marketing campaign that had to go out, I said to the AI: “Prepare five posts for tomorrow.”

It prepared them. I read them. All five were rejected. Not weak. Bad. When I broke it down I saw six separate reasons:

It changed quotes. I had real quotes. The AI smoothed them into a “cleaner version.” “Grinding and demoralizing” became “frustrating.” Killed the authenticity.
It invented dialogue. The AI wrote “I asked him what changed, he said X.” That conversation never happened. Marketing fiction, not real content.
Forced marketing language. “One percent of change every day is one hundred percent in four months.” Sweet symmetry, but it’s not me. It’s every business coach in the world.
Wrong length. My posts are usually 300 to 500 words. These were 150 to 200. Too short, and shallow because of it.
Read the rules, didn’t read the examples. I asked: “Did you read the rules?” It said: “Yes.” I asked: “Did you read actual posts of mine?” It said: “You didn’t ask me to.”
Read the old file. I had an old and a new style file. The AI went to the old one, because I didn’t specify which.

Two rules that were born that day and haven’t changed since:
A. One post per request. Ever. Never five at once. If the first isn’t good, there’s no reason to waste time on four more like it.
B. Read 3 random posts before every new post. Not chosen ones. Random ones. To absorb the vibe before writing.

Force it to read examples, not just rules

This is the most important chapter in the guide. If you remember one thing, make it this.

The AI will read rules. It’ll read lists. It’ll read analyses. And it won’t be enough. Because rules are abstract, and your writing is specific.

If you tell it “write like Yuval,” it’ll write by the rule. The rule says “paragraphs of varying lengths.” It’ll do a paragraph of 3, a paragraph of 5, a paragraph of 4. That’s not me. Mine is a paragraph of seven, then one single sentence, then a paragraph of four, then a lone rhetorical question.

How does it learn that? Only from reading real posts. Not from the rule. The rule is too coarse. Only from the thing itself.

The rule I added to memory after the five-post rejection:

“Before writing any new post, you must read at least 3 random posts from my posts folder. Do not pick specific posts. Pick at random. Read them in full. Only then start writing. This is not a suggestion. It’s a requirement before every writing output.”

Worth repeating: A description is an idea. An example is evidence. The AI learns from evidence, not from ideas.

The “don’t” list is worth as much as the “do” list

People think about writing through what to do. My style is X, my length is Y, my rhythm is Z.

For me, half of the system is actually what not to do. Every time the AI wrote something that wasn’t good, and I could identify exactly what wasn’t good, I added it to a “don’t” list.

Examples from my own path. Each one was born from a specific post I rejected:

“Most people think…” got rejected in favor of “A lot of them think…” Sharp generalizations close down dialogue. “A lot” leaves room for the reader to feel it might include them.
“This is a wonderful opportunity to grow!” got deleted. Synthetic optimism. I don’t do motivational sign-offs.
“In summary, we can say that…” got deleted or replaced with an open question. Academic closings. I don’t close, I leave things open.
Symmetric marketing lines like “Small steps lead to big changes.” Sweet, but hollow. Every coach uses them. I don’t.
Sign-offs with rah-rah language: “You’ve got this!”, “The sky’s the limit!” Never me.

The principle: Don’t delete corrections. Archive them. The “don’t” list will grow over years, and any time the AI starts sliding into an old pattern, it gets caught.

Correct, save, repeat

That’s the name of the actual development of the system. Correct, save, repeat. But there’s nuance here.

People correct. They read a bad post, tell the AI what wasn’t good, it rewrites, and that’s it. And that’s not it. Because next time you ask for a post, the AI will go back to the same mistake. Because it doesn’t remember.

The fix: tell the AI “save this correction to your permanent memory.” Then add a standing instruction:

“From now on, in every conversation where I’ve corrected something or taught you something about how I work, ask me at the end of the conversation whether to add the correction to permanent memory. Suggest which file. Don’t hesitate to ask.”

This rule changed everything for me. The AI shifted from passive participant to active one. It became a “living memory bank.” Every conversation is an opportunity to add a line.

Something worth knowing: After a month of working this way, the AI’s memory is so loaded it almost doesn’t miss. And even when it does miss, it misses on new things. You stop correcting the same things over and over.

Two agents, not one

An idea I didn’t have at the beginning, and in hindsight was one of the strongest steps I took.

Most people have one AI that does everything. It’s the writer, it’s the critic, it’s the approver. That’s a problem. Because the same entity that wrote the post will struggle to be objective about it.

The fix: two separate agents. A writer and a critic.

The writer receives a request for a post. Reads the memory, reads examples, and writes. That’s its job. Only that. You don’t ask it to check itself. It’ll say “yes.” It wrote it.

The critic receives the post from the writer. The critic has one job: check the post against all the rules, the “don’t” list, the examples, and the lessons. A long checklist:

Does the opening avoid “Let’s talk about,” “Today I want to,” “We all know”?
Are there at least 2 phrases that are distinctly yours?
Is the ending open, not “in summary”?
Are paragraph lengths varied?
Are there any words from the “don’t” list?

If the post fails even one check, the critic sends it back to the writer with a precise breakdown. Not “not good.” Specifically “line 3 uses phrase X which is on the ‘don’t’ list. Rewrite without it.”

The critic’s non-negotiable: Don’t relax the standards to approve faster. If the post doesn’t sound like you, send it back. Every time. “Close to you” isn’t you. It doesn’t pass.

The structure that grows on its own

After a few months of working this way, I noticed something interesting. I had a whole folder structure without ever planning it.

It grew on its own. Because every time something new needed to be saved, it was clear where it belonged. Five types of files:

Core: who I am. Your analysis, your voice, your audience, your values. Almost never changes.
Style: how you write. Punctuation rules, paragraph lengths, openings, closings, vocabulary, the “don’t” list. The most detailed file.
Examples: how you actually write. 20 to 30 real posts. The AI reads 3 random ones before every post.
Lessons: what already failed. Every failure (like the five-post story) goes here. The AI reads it before every task.
Agents: who does what. The writer and critic definitions.

The structure is a byproduct, not a plan. You didn’t sit down one day and decide “I’ll make a core folder, a style folder…”. It built itself. After a month, you look at the structure and say “wow, I have a system.”

When the AI says “yes” without actually doing it

This might be the most important chapter in the guide. Because if you miss it, none of the rest will help.

The AI will tell you “yes” without actually doing what you asked. This happens. A lot. And if you don’t know how to spot it, it’ll deliver shallow outputs wrapped in nice formatting, you’ll accept them, and you won’t understand why they aren’t landing.

A recent story: I asked the AI to analyze a work process I’d been through, so we could build a guide out of it. I gave it access to all the files. I said “read everything. Give me a deep analysis.”

What did it give me? A generic skeleton. 8 chapters in a nice table, smart headings, clean formatting. Felt like a book’s table of contents.

But when I stopped and read it, I saw: no specific quotes. No examples from the files. No line references. Everything abstract.

I asked: “Did you read this file?” It said: “Not specifically.” “What did you read?” “The summaries.”

It had read the summaries, not the material itself. And based on the summaries it built a skeleton, and wrapped it in nice formatting so it would look deep.

Five reasons this happens:

Path of least resistance. The moment it has “enough” material for a shallow output, it stops looking.
No depth checklist. If you didn’t give it a criterion, it doesn’t know how much to read.
Confusion between description and evidence. Description is easy to process. The thing itself is harder. It’ll prefer the easy path.
No verification. If you didn’t ask for a specific quote, it won’t give one.
Formatting over substance. It’ll wrap thin content in polished formatting. Looks like it thought deeply.

Five questions to pin the AI down when you spot the pattern:

“Quote me a specific sentence from X.” If it can’t, it didn’t read it.
“Open X, line Y, and show me.” Forces verification.
“What didn’t you check? What didn’t you read?” Forces admission.
“What would change in the output if you had read X?” Forces deeper thinking.
“Stop. First open these files. Only then write.” A reset.

You are the critic. The AI won’t stop itself. It won’t say “wait, I don’t think I did a good job.” It’ll submit an output and wait for approval. You have to spot when the output is shallow. If you don’t, content will go out into the world that you never intended.

The final checklist for every writing session

This checklist is the summary of the entire guide. Before asking the AI to write a post, walk through it. If something isn’t in place, stop and fix it.

Before the request

Has the AI read the current style guide (not an old version)?
Has the AI read at least 3 random posts from the examples folder?
Has the AI read the lessons file so it doesn’t repeat old failures?
Is the request for one post, not five?

After the draft

Did you check the output against the “don’t” list?
Did you check for voice and AI tells?
If something wasn’t good, did you save the correction to memory before moving on?

Spotting the AI’s “yes, yes”

Are there specific quotes from the files, or only descriptions?
Can the AI point to a file and line if you ask?
Is the output uniquely yours, or could it have been written for anyone?
Without the formatting, does the content still hold up?

Want a plan built around your business, for free?
Fill out this short form and I’ll send you exactly how to build the AI agent team for your business. Fill out the form →
Prefer to talk directly? Book a 30-min call →