Blind Dev article

How to gradually teach an AI to write in your style

A process for examples, feedback, voice, quality control, and avoiding generic output.

Based on Blind Dev personal experience and source posts; verify current details before applying.

Archive publication: the results, tool limitations and plans described here reflect the time of the original text.

I had a recurring problem: AI helped me write, but often made the text too polished, formulaic and unlike my own writing.

It often avoided expressing an opinion, added unnecessary words and generally wrote too correctly. Most importantly, the texts had almost no emotion.

I wanted something different: help shaping my own thoughts in my own style, with simple language, a clear structure and less unnecessary AI-generated material.

I did not look for one “magic prompt”

There isn’t one.

Instead, I created a separate ChatGPT conversation specifically to identify my style.

In the first message, I assigned the AI its role and working rules:

In this chat, we are defining my style. I will record voice messages. Your task is to turn them into text without stumbles, using simple words and a clear structure. Do not write in a formulaic, overly polished, emotionless way (if I spoke with emotion), or use bureaucratic language. If necessary, make an outline first, then write the text. I will point out phrases that do not sound like me, and you should take that into account going forward. Your task is to gradually adapt to the way I write.

To start recording, I click “Start dictation.” Voice mode is another option, but ChatGPT will occasionally speak if you leave long pauses.

Here is how I worked after that

• I dictated thoughts on different subjects.
• I asked the AI to understand the meaning first and make an outline if necessary.
• Then I asked it to turn that into a text.
• If the result did not sound like me, I did not just ask for a rewrite. I explained exactly what was wrong.

For example:

Too polished. I would write it like this: “…”

Too formulaic. My version: “…”

I do not talk like that. Here is how I corrected it: “…”

Too many connecting phrases. Just get to the point.

Too correct. It needs more life and the words I actually use, such as “well,” “honestly” and “it annoys me.”

Gradually, the AI began to understand what I wanted better.

Then I realized something else

It is better to save these rules separately in a Markdown file so you can use them in new chats.

Prompt:

Save my style in a .md file that I can send in a new chat so you can keep writing posts with no corrections or as few as possible. In that chat, you will also update it if I give you more corrections.

It is convenient. You do not have to explain everything again. Give it the file and keep writing.

But after it provides the file, I recommend asking:

Did you include all the rules? Did you leave anything out?

One prompt is not enough

If you simply say “write like a human, without AI-generated phrasing,” the AI will still drift back to overly correct, formulaic text.

What worked for me was a dedicated style chat, several texts on different subjects, specific corrections, and then saving the rules in a Markdown file.

The results became noticeably closer to my own style: fewer stock phrases, less dryness, and fewer obvious signs of AI writing.

But it still does not become perfect. If I stop editing, the AI starts writing too smoothly again. This is not something you set up once and keep forever.

One person I know spent almost a month shaping their style through voice messages!

You can also provide an archive of old posts

Use a file containing your texts, or a list of links if the model can open and read them. A file is more reliable, though.

Save several texts, around 10–20, in one file and ask it to identify the features of your style.

Prompt:

Look at the attached file: it is a collection of posts from my Telegram channel. Based on how I wrote those posts, is there anything to add to the style file you created? Anything to change or clarify? First write a list; I will tell you whether to make the changes.

The original article includes a screenshot of the model’s response. Then check how well those rules actually match the way you write.

ChatGPT sent confusing code in the chat. What now?

Just say:

Give me a download link. That is easier.

In my case, I asked it to add everything except items 4 and 5 from its list of proposed additions based on my posts, and then provide a link.

Always check the file by opening it in an editor such as Notepad, VS Code or Notepad++.

For some reason, the AI shortened mine.

I told it:

You sent 2,098 characters in this file, while the previous one had more than 4,000. Did you delete something? I asked you to add the relevant points from the posts, not cut things. Here is the old version again: …

In place of the ellipsis, I pasted the code GPT had sent earlier.

It gave me a proper file. But when I asked for a download link, it gave me a shortened version again.

Still, I think it is easy enough to press “Copy,” paste into an editor, choose the .md file type and save.

The result

If you also use AI to write posts, articles or replies, this approach can genuinely help.

It does not guarantee a text that sounds 100% human. But it makes editing much easier and lets you write even while on the move, just by dictating voice messages!

Case study author: Denis Skripnik.

Source: full Russian article on VK; author’s announcement. English translation. Conversation screenshots are available in the original.