Describe My Writing Tone to AI? Adjectives Are Not Your Voice

TL;DR: If you want to describe my writing tone to AI and get something back that sounds like you, adjectives will not do it. "Professional yet friendly" describes thousands of people, so the model returns the average of all of them. AI does not know your tone; it knows words that describe tone. A voice lives in sentence length, qualifier placement, transitions and punctuation, and every one of those is countable. Measure them from your own writing, state them as instructions, and paste the result into ChatGPT, Claude or Gemini. That is the difference between a tone label and a Style Profile.

You built a Custom GPT called "Write Like Me". The instructions say: warm but professional, concise, confident, no fluff. The first draft it produced opened with "I hope this message finds you well" and closed with "Please don't hesitate to reach out." You rewrote the instructions twice. The third version sounds exactly like the first.

The instructions were followed. The problem is what they contained.

Can You Describe Your Writing Tone to AI With Adjectives?

You can type them, and the model will comply, and the result will not be your voice. An adjective is a pointer to a register the model already knows. It says nothing about which of the thousands of writers inside that register you are.

Consider what the model has when you write "professional yet friendly" into custom instructions. During training it saw an enormous amount of text labelled or described with those words: corporate blog posts, HR emails, customer-support replies, LinkedIn updates. The phrase now points to the statistical centre of all that text. When you invoke it, you get that centre. Your colleague who types the same phrase gets the same centre. The two of you write nothing alike in real life, and the model has no way to know that, because the only thing either of you gave it was the label.

Run the experiment yourself. Put "professional yet friendly, concise and warm" into a fresh chat and ask for a follow-up email after a client meeting. Ask someone on your team to do the same. The two drafts will resemble each other more than either resembles anything you actually sent last month. No amount of adjective stacking changes that. "Confident, direct, approachable, human" is four pointers to four registers, and the model blends them into one average.

We have measured how narrow that average is. In our data on why AI writing sounds generic, we scored 320 AI samples across five models on six style dimensions, and the spread between models was small next to the distance between any model and an individual writer. The tone label does not move the model off its default. Nothing in the label tells it where to go.

Why Does "Professional Yet Friendly" Produce the Same Email for Everyone?

Because AI does not know your tone; it knows words that describe tone. A tone label is a description of an effect, and the model has no record of how you produce that effect.

Think about what "friendly" covers. One person is friendly by opening with a personal line before the ask. Another is friendly by using first names and short sentences. A third is friendly by adding a light joke at the end and never in the middle. All three would tick "friendly" on a form. All three write completely different emails. The word collapses those differences into one register, and the model can only produce the register.

The same collapse happens with every adjective on every tone template on the web:

Every row ends the same way. The label names the effect, and the model fills in the mechanics with its default. OpenAI's own guide to custom instructions describes the field as a place to tell ChatGPT how you want it to respond. Most people fill it with effects. The model needs mechanics.

This is also why your third Custom GPT sounds like your first. A Custom GPT is a container. Rewriting the adjectives inside it rearranges pointers to the same average. Our guide to the best ChatGPT custom instructions shows what those fields look like when they carry mechanics instead of effects.

What Is a Voice Actually Made Of?

A voice lives in sentence length, qualifier placement, transitions and punctuation. Those four are where readers recognise you, and none of them is an adjective. Each one is a measurement you can take from writing you have already done.

Sentence length. Not "concise" but a distribution. Your typical sentence might run twelve words. Your short ones might drop to four when you are making a decision. Your long ones might reach thirty when you are laying out context. The rhythm of switching between them is yours, and a model given "concise" will produce uniform fourteen-word sentences forever.

Qualifier placement. Whether "probably" opens the clause or closes it. Whether you hedge before a recommendation and never after one. Whether you use "I think" at all once you have decided. Two writers with identical vocabularies differ here on every paragraph, and readers feel it without being able to name it.

Transitions. "So" versus "therefore" versus a bare new sentence. Whether your paragraphs open with the conclusion or build to it. Whether you use "and" to start a sentence. Whether you ever write "however". The model's default transitions are the ones it saw most, and they are the ones that make a draft smell like AI.

Punctuation. Colon frequency. Comma density per sentence. Whether asides go in parentheses or get their own sentence. Whether you use exclamation marks in professional email (some people do, once per message, in the same place every time).

On top of those four sit vocabulary and avoidance: the words you reach for and the phrases you would never write. The point is the same in every case. These are countable. You can take ten emails you sent and produce a number for each one. A number can be stated as an instruction, and an instruction with a number in it can be checked. Our explanation of how style extraction works walks through the full set of dimensions we measure, and there are more than four, but these four are the ones that separate a tone label from a voice.

Why Do Samples and Custom GPTs Fall Short Too?

Samples beat labels and still leave the model inferring. Custom GPTs are containers and inherit whatever weakness sits inside them. Neither fixes the underlying gap on its own, which is that nobody has told the model what your patterns are.

Pasting three emails into a Custom GPT's instructions, or uploading them as knowledge files, gives the model raw evidence. That is a real improvement over adjectives. The model can now see that your sentences are short and your asks come first. What it has to do next is guess which of the patterns in those three emails are yours and which belong to that one Tuesday. Three samples cannot separate a habit from a mood. The inference also drifts. A long conversation pulls the model back toward its default with every turn, because the samples sit at the top of the context and the model's own output accumulates below them.

There is a practical ceiling as well. The custom instructions field has a size, and samples consume it fast. Fill it with three emails and there is no room for the rules you actually need. Move the samples into knowledge files and the model consults them when it decides they are relevant, which is not on every request.

Here is the pattern across the three approaches most people try:

The container question (custom instructions, Custom GPT, Project, Gem) matters far less than what goes into it. We compared what ChatGPT's memory actually stores against a measured profile in our memory versus Style Profiles comparison, and the conclusion holds for every container: memory stores facts about you, and facts about you are not your voice either.

How Should You Describe Your Writing Tone to AI Instead?

Measure it from your own writing, state the measurements as instructions, and paste those instructions into whichever container your AI offers. Replace every adjective with a number or a rule the model can follow without interpreting.

The manual version takes an hour and beats any tone template:

1. Collect ten to fifteen pieces you wrote and were happy with, in the register you need most. Client emails, internal updates, proposals. Pick the ones people replied to well. Our guide to selecting writing samples covers what makes a sample useful and what to leave out.
2. Count words per sentence across all of them. Note the typical value, the shortest you go, the longest you go. Write it down as a range.
3. List every hedge you used and where it sat. Beginning of the clause, end of the clause, or absent after a decision.
4. List your transitions. The ones you used, and the ones you never used. The second list is often more useful than the first.
5. Note your punctuation habits. Colons per email, commas per sentence, parentheses or separate sentences for asides.
6. Write the findings as instructions. "Sentences average 12 words, range 4 to 28. Open with the ask. Hedge only before recommendations, never after decisions. Never use 'however' or 'moreover'. One colon per email at most." No adjectives anywhere.

That block goes into ChatGPT custom instructions, a Custom GPT's instructions field, a Claude Project's system prompt, or Gemini's saved info. Same block, every tool.

This is what a Style Profile from MyWritingTwin.com is: the measurement step done for you, across more dimensions than you would count by hand, from samples you already have. The output is a Writing DNA analysis expressed as constraints, packaged as a Master Prompt you deploy once inside the AI you already pay for. Your Writing Twin for ChatGPT, Claude and Gemini is that block of instructions, not a separate chatbot, and it lives inside the container you already use.

Two properties of a measured profile matter for the argument here.

It can be checked. "Was that friendly enough?" has no answer. "Was the mean sentence length within the range?" does. When a draft drifts, you count, you see where it missed, and you correct one rule. An adjective gives you nothing to correct.

It survives the next model release. The measurements are about you, so they do not change when the model does. A better model executes the same profile more faithfully than the last one did, which is why a measured profile becomes more useful with each release while a tone template stays exactly as vague as the day you wrote it.

How Do You Check Whether Your Tone Description Worked?

Give the model the brief for an email you already sent, then compare its draft to yours by counting. Five minutes, no tools, and the result is a number rather than a feeling.

Pick an email from last week. Write a two-line brief: who it is to, what you need, what they already know. Paste the brief into your assistant with your current tone instructions active and ask for the email. Put its draft next to yours and count:

• Words per sentence, shortest and longest
• Sentences per paragraph
• Where the ask sits (first line or last)
• Hedges: how many, and whether they open or close the clause
• Commas per sentence
• Opening line and closing line

If the numbers differ, your description did not work, whatever the assistant tells you when you ask it directly. If they match on one email, try three more in different registers. Tone adjectives usually survive the first test and fail the second, because the label was written for one mood and you have several.

Run the same test after you replace the adjectives with measurements. The numbers converge. That convergence is the whole point, and it is what "describe my writing tone to AI" should have meant from the start: not a description of the effect you hope for, but a specification of how you actually write.

Your voice was never an adjective. It was always a set of habits you could count. Once you hand the model the count, it stops averaging you with everyone else.

See how far your current setup is from your real writing in about a minute with the Humanity Score test, or build a Style Profile from your own samples and paste it into the AI you already use.