---

An adjective is not an instruction. "Human" carries no clear meaning for the model, so it falls back on its average: even sentence lengths, lists of three, transitional filler. What you need are rules, modules and your own sentences.

Level 1: The System Prompt

This is no place for adjectives. It is a place for rules you can verify. The difference is simple: the model cannot measure "write naturally", but it can measure "one subordinate clause per sentence, maximum".

The rule of thumb: however you would say it out loud, write that down once and put it in.

My block looks roughly like this:

Writing rules:
- Second person, throughout.
- One thought per line. No nested clauses.
- No lists of three in running prose.
- Vary sentence length: every paragraph needs at least one
  sentence under 5 words and one over 20.
- No em dash. Use a comma, period, colon or parenthesis.
- No transitional filler (furthermore, essentially,
  it is worth noting).
- Every claim gets an example or a number.

Verifiability is the whole point. You can count every one of those lines on a finished text with your finger. If you cannot count a rule, rewrite it.

If you work with Claude Code or Cowork, this block belongs in your CLAUDE.md. How that file is structured: Setting up CLAUDE.md.

Level 2: Skills Instead of One Giant Prompt

At some point the single prompt gets too long and the model starts ignoring parts of it. That is when you split it into modules: one for tone, one for style, one for structure, one for format. Each module fires only when it is needed.

The most valuable module is one that checks instead of writes. I built a skill that scans my texts for AI patterns:

• cadence and uniform sentence lengths
• three-beat rhythm in prose
• repeated words inside the same paragraph
• hidden Unicode characters that ride along when you copy out of AI tools

A skill that is just called "humanizer" and tells the model to write more human changes nothing at all. That is level 1 in a different folder. What you need is a linter: a module with fixed check rules that hands you findings with line numbers.

You can grab mine: Anti-AI Writer. Download and setup for Cowork, Claude Code and standalone Python are all there. If you would rather build your own, start with Claude Skillcreator.

Level 3: Persona Vectors

Now the uncomfortable part. There is a layer beneath prompts and skills, and you cannot reach it.

In 2025 Anthropic showed that a model's character traits can be located as activation patterns inside the network. They call these persona vectors and use them to measure and steer personality shifts during training and deployment (Anthropic Research, July 2025, arXiv 2507.21509).

That happens at the level of the weights. Shifting anything there would require access to the model itself, which you do not have with Claude or ChatGPT. So when someone sells you the idea that your model will eventually sound like you on its own, given enough training: that is not how the system works.

What you can do instead is cheaper and works immediately. You supply your voice as an example rather than as a description.

Here are five paragraphs I wrote myself. Analyse my sentence
lengths, my rhythm, my typical transitions and the words I
reach for unusually often. Do not summarise the analysis for
me, just work in that pattern from now on.

[your 5 paragraphs]

Five paragraphs are usually enough. They have to be genuinely yours, otherwise you are training the model on a voice that is not. If you have little written material, take voice notes you sent to friends and have them transcribed. Those sit closer to how you actually talk than anything you ever wrote for a client.

Common Mistakes

• Adjectives as rules. "Natural", "casual", "authentic" are not measurable for the model.
• Everything in one prompt. Past a certain length something at the end drops out, and you will not notice which part.
• Not proofreading the rewrite. The model likes to smooth text into a fresh pattern that is just as uniform as the old one.
• Sample texts from other people. You get somebody else's voice, cleanly reproduced.
• Chasing a detector score. AI detectors measure perplexity and burstiness. A good score does not mean the text sounds like you.

Next Steps

1. Write down your seven rules. Only the ones you can count.
2. Take a finished text of yours and count how often you break your own rules. That sharpens the list faster than any example.
3. Install a linter, mine or your own.
4. Put five of your own paragraphs into a file you carry into every new project.