You Do Not Know How You Write, So Quit Pretending Like You Do

Your AI voice profile is sitting in your chat history. Stop building it from scratch. Start mining the archive of you.

You think you know your writing voice. You don't.

Ask anyone who writes for a living to describe how they write, and you'll get a clean little answer. "I'm direct." "I'm warm." "I like short sentences." Cool. Now go read their last 50 messages and tell me if any of that is true.

I'll wait.

The truth is that most of us are walking around with a sanitized, marketing-deck version of our own voice in our heads. The version we want to be true. The version we'd write on an About page. But the actual way we write, the way we sound when no one is grading us, lives somewhere we never look.

I'm going to tell you where it lives. And how to get it out.

The Popular Advice (And Why It Almost Works)

If you've been anywhere near AI content lately, you've probably seen the advice: build an "about-me" file. Build a "voice profile." Build an "anti-AI-writing-style" file. The gurus will tell you to do this so that every time you ask Claude or ChatGPT to write for you, it sounds like you, not like the smooth, hedge-everything, em-dash-loving robot it defaults to.

And honestly? They're right. You should have these files. They are the difference between sounding like Claude and sounding like yourself.

But here's where the advice goes sideways.

The popular method tells you to sit down with Claude and run a 100-question self-interview. Answer methodically. Describe your tone. Describe your audience. List your non-negotiables. Build the file from scratch.

Meanwhile.

Your chat history is sitting right there.

You've been talking to ChatGPT or Claude for months. You've told it things. Lots of things. Relationship drama at midnight. Frustrations about a client who didn't pay on time. The half-baked business idea you typed out before you went to sleep. Complaints about your mother-in-law. Questions about whether you should leave your job. ChatGPT has seen things.

Why are we acting like all that raw material doesn't exist? Why are we sitting down to construct a voice profile from scratch when there is already an archive of you, written by you, time-stamped, organized by chat title, sitting in your account right now?

The raw material already exists. You just never thought of it as a voice database.

Describing How You Write vs. Showing How You Write

Here is the flaw with the interview method.

Answering interview questions is performing self-knowledge. You sit up straighter when someone asks you to describe your tone. You overthink it. You give the answer you wish were true. You say "I'm direct and warm" because that's what people who wish they were direct and warm say about themselves.

Your chat history is evidence of how you actually write.

There is a difference between describing how you write and showing how you write. One is constructed in real time, while you're trying to be coherent and self-aware and impressive. The other was captured in real time while you were trying to be none of those things. You were just trying to get help. You were just trying to vent. You were just trying to figure out how to phrase the awkward Slack message to your boss.

That is the version we want. Not the curated one. The captured one.

Showing beats describing, every time.

What This Has To Do With LLM Instance Cloning

I've written about LLM Instance Cloning before. If you haven't read it, here's the quick version, because the rest of this post won't make sense without it.

LLM Instance Cloning is the process of asking an AI to document its own behavior patterns, preferences, and decision-making processes within a specific conversation, then packaging those insights into a reusable prompt that recreates the same "personality" in a new conversation.

When I first wrote about this, I was using it to clone the AI. I had spent weeks training a Claude Instance to refine my blog posts exactly how I wanted, and I needed to capture that trained behavior before the chat ran out of space. So I asked Claude to write down what it had learned about helping me, and I used that document to spin up new Claude Instances that already knew how I liked things done.

This post is a different flavor of the same technique.

Instead of cloning what the AI learned about helping you, you're cloning what YOU revealed while it was helping you.

The AI was never the one being trained in this scenario. You were the one leaving breadcrumbs. The technique just makes you go collect them.

The Actual Method

Here's how I do this.

Step 1: Pick Your Chats

Go into your AI of choice (Claude, ChatGPT, whatever you ramble in the most) and pick three to five chats. Not your to-do list chats. Not the chats where you asked it to help you fix a CSV. The chats where you were the most YOU.

You know the ones. The chats where you complained. The chats where you vented. The chats where you brainstormed at 11pm and the messages got progressively less professional as your eyes got heavier. The chats where you talked about your business, your work, your people, your fears.

Those.

If you're not sure which to pick, scroll through your chat history and look for the long ones. Length is usually a tell. If you wrote a lot, you were probably also being yourself.

Step 2: Extract an About-Me File

Go to the END of one of those chats. Don't start a new chat. Stay in the chat that has the evidence. Then paste this prompt:

Review the entire conversation we have had in this chat from the very beginning.

Based on EVERYTHING I have said (not what I have asked you to do, but what I have revealed about myself), compile a structured About-Me file in markdown.

Include the following sections:

## Who I Am
A short biographical paragraph based on what I have shared about my background, role, work, and life.

## What I Care About
The values, themes, and topics that come up repeatedly across this conversation.

## What I Am Working On
The projects, goals, and current focus areas I have referenced.

## How I Think About My Work
The way I describe my own work, what I am proud of, what frustrates me, what motivates me.

## Patterns and Recurring Concerns
Things I bring up more than once. Worries. Obsessions. Things I cannot let go of.

Pull direct quotes from the chat wherever possible. Do not paraphrase what I said unless you have to. The whole point is to capture the actual me, not a polished version.

Output as a downloadable markdown file.

Run this in each of the three to five chats you picked. You'll end up with three to five raw About-Me files, each one shaped by the kind of conversation it came from.

Step 3: Extract a Voice Profile

Same chats. Different prompt. Run this one next:

Review the entire conversation we have had in this chat.

I am building a Voice Profile that will be used as a reference for AI to write in my voice. Based on how I have written in this chat (not the topics, the WRITING), compile a structured Voice Profile in markdown.

Include the following sections:

## Rhythm
Average sentence length. Do I write in fragments? Long flowing sentences? Short punchy ones? Mixed?

## Openings
How do I tend to open messages? Do I dive in? Do I ramble first? Do I greet?

## Closings
How do I tend to close messages? Do I sign off? Trail off? Pose a question?

## Asides and Jokes
Do I make jokes? What kind? Do I use parentheticals? Do I interrupt my own thoughts?

## Vocabulary
Words I use often. Words I repeat. Filler words. Intensifiers.

## What Makes My Writing Recognizable
If you read a message in isolation with no name attached, what would tell you it was me?

Pull direct quotes from the chat as evidence for each section. The output should feel like a forensic report on how I write, not a flattering description.

Output as a downloadable markdown file.

This is the one that surprises people the most. The voice profile that comes back is rarely the voice you THINK you have. It's usually closer to your friend group's description of you than your LinkedIn bio.

Step 4: Extract an Anti-Writing-Style File

Same chats. Different angle. This one captures the negative space, what you reject.

Review the entire conversation we have had in this chat.

I am building an Anti-Writing-Style file. The goal is to document everything I have pushed back on, edited away, or expressed dislike for in how AI writes.

Include the following sections:

## Words and Phrases I Have Pushed Back On
Anything I have asked you to remove, replace, or stop using.

## Tones I Have Rejected
Moments where I said something sounded corporate, formal, generic, robotic, or off.

## Structural Choices I Have Rejected
Lists I asked to be turned into prose. Headers I removed. Em dashes I asked to delete. Bold I told you to strip out.

## Emotional Tone I Avoid
Moments where the writing felt too earnest, too performative, too hedged, too anything.

## What Makes Me Cringe
If there are patterns in what I have rejected, name them. Cluster the cringes.

Pull direct quotes wherever possible. Show the moments where I said no.

Output as a downloadable markdown file.

This file is the one most people skip when they build context layers from scratch. They build the positive picture (here's how I write) but never the negative one (here's how I never write). The negative one is the one that prevents AI from defaulting back to its smoothed-out, jargon-friendly, em-dash-everywhere baseline.

Step 5 (Optional): Extract a Beliefs and Positions File

If you've been ranting in your chats, opining about your industry, picking fights with ideas you disagree with, declaring opinions you'd defend at a dinner party, run this one too:

Review the entire conversation we have had in this chat.

I am building a Beliefs and Positions file. The goal is to capture the opinions, hot takes, and recurring positions I hold.

Include the following sections:

## Things I Strongly Believe
Statements I have made with conviction. Claims I have defended.

## Things I Reject or Push Back On
Ideas, trends, or framings I have called out as wrong, bad, or overrated.

## Hills I Would Die On
The opinions that come up more than once. The things I clearly will not be talked out of.

## What I Am Skeptical About
Where I have expressed doubt or refused to accept the consensus view.

Pull direct quotes as evidence. Do not soften my positions. If I said something with conviction, keep the conviction.

Output as a downloadable markdown file.

This file is gold for anyone writing thought-leadership content, op-eds, LinkedIn posts that take a stand, or any writing where a point of view is the whole point.

The Consolidation Step

After all that, you'll have three to five sets of files. One set per chat. They will overlap. They will also contradict each other in small ways, because you're not the same person in every chat.

Now open a brand new chat. Upload all of them. And paste this:

I am building a final set of context files that capture who I am as a writer and a person. I have attached extractions from multiple chats in which I have been myself in different moods, on different days, talking about different things.

Synthesize all of these into a clean, final set of files:

1. About-Me.md — a unified picture of who I am, what I care about, what I am working on, what I think about my work.
2. Voice-Profile.md — a forensic description of how I write, with examples.
3. Anti-Writing-Style.md — everything I have pushed back on, every cringe, every rejection.
4. Beliefs-and-Positions.md (only if relevant material exists) — the recurring opinions and hot takes that show up across multiple chats.

For each file:
- Pull direct quotes from the source extractions as evidence.
- Where the source files disagree, note the disagreement and pick the version that shows up most often.
- Keep each file under 2,000 words. If you have to cut, cut the flattering parts. Keep the specific, the weird, the inconvenient.
- Output each file separately as a downloadable markdown artifact.

What you get back is a clean set of files that were not invented in a 100-question interview. They were captured from the archive of you, then distilled.

What You Do With Them

These files become your context layer.

Every time you write something, every time you ask an AI to write for you, every time you sit down to draft an email or a blog post or a LinkedIn caption and you don't feel like building a fresh prompt from scratch, these files go in first.

Upload them to your AI's Project. Pin them in your Cowork folder. Save them as a Project file attachment in Claude. Wherever they live, the rule is the same: the AI reads them BEFORE it writes for you.

Claude sounds like you. Not like Claude.

This is what the gurus are actually trying to get you to build. The end goal of the 100-question interview is to produce these exact files. You just got there faster, with more honest material, by mining the archive you already had.

If You Want Help With This

I build custom AI solutions, and part of that involves building writing systems. Systems that take you from rough draft to publishable in minutes, not hours.

The writing systems I build do not replace your voice. Think of it like a system that takes a rough sketch and turns it into a polished painting, except the painting still looks exactly like the artist intended. The brushstrokes are yours. The colors are yours. The system just handles the part where you sit there for three hours staring at a blank canvas.

This matters to me. I am not in the business of making you sound like someone else. I am in the business of making your own voice ship faster.

Three kinds of people benefit from this right now:

• Ghostwriters who want a system to manage multiple clients without their voice profiles bleeding into each other. One context layer per client. Every draft sounds like that client, not like your previous client.

• Business owners who know they should have started writing on social media last year and are finally ready to start today. You have the expertise. You have the opinions. You do not have a thousand hours to build a writing habit from scratch. A system collapses that timeline.

• Anyone who reads this and realizes they do not want a handful of prompts, they want the actual system underneath. The context layer in this post is step one. The full build is agent definitions with multi-step workflows, automated queues that run on a schedule, knowledge bases structured as raw data the AI can query, and MCP-level infrastructure wired into the tools you already use. If you want the foundation, the prompts above are enough. If you want the ceiling, that is the conversation I am here for.

I build systems for blogs, social media, newsletters, whatever channel you’re using to put your work in front of people.

If that sounds like what you need, reach out to me here. Tell me what you’re working on. I’ll tell you what a writing system would look like for it.

As always, thanks for reading.