Custom Instructions vs Style Profile: Which One Sounds Like You?

TL;DR: Custom instructions vs style profile is a question of source, not of format. Custom instructions are a description: what you tell ChatGPT about how you want to sound. A Style Profile is evidence: measurements taken from writing you already did, covering sentence length, qualifier placement, transitions and punctuation. A prompt describes how you hope to sound; a profile built from your own writing shows how you actually write. The two are layers rather than rivals, because the profile is deployed through the custom instructions field. The question is what goes in the box. Below: why self-described instructions plateau after the third rewrite, what a measured profile contains that a description cannot, and how to switch this week.

You pay for ChatGPT Plus. You have rewritten your custom instructions three times, built a "write like me" Custom GPT, and you still rewrite every draft before it goes out. The instructions are not the problem. The source of the instructions is.

What Is the Difference Between Custom Instructions and a Style Profile?

Custom instructions are a settings field where you describe how you want the model to respond. A Style Profile is a document produced by analysing writing you already did, then written as instructions the model can follow. The first is a self-report. The second is a measurement.

OpenAI documents the field in its custom instructions guide: one box about you, one box about how responses should read, applied to every new chat. The container is fine. Roughly 1,500 characters for the first box and up to 8,000 for the second, per our character-limit breakdown, is plenty of room for a voice specification.

The problem is what people put in the box. Ask a professional to describe their writing and you get adjectives: direct, warm, concise, professional but approachable. Those words are honest, and they are useless as instructions, because they describe a register shared by everyone who chose the same adjectives. The model has read millions of documents labelled "professional" and returns their average.

A Style Profile skips the self-report. MyWritingTwin.com takes three to five pieces you wrote yourself and extracts the Writing DNA: sentence-length range and variation, where your qualifiers sit, which transitions you use and which you never use, punctuation density, opening and closing moves, and the phrases that never appear in your writing. The output is a Master Prompt, plus a condensed Runtime Block sized for fields with character limits. The Runtime Block goes into the same custom instructions box. Same container, different contents.

Why Do Custom Instructions Stop Working After the Third Rewrite?

They stop working because every rewrite improves the description, and a description has a ceiling: your own knowledge of how you write. Most people cannot state their average sentence length, how often they hedge, or which connective they reach for when they change subject. The third rewrite is usually as good as self-report gets.

Here is the loop most ChatGPT Plus users recognise. Version one says "write in a professional, friendly tone and keep it concise." The output reads like a competent stranger. Version two adds rules: no bullet points in emails, no "I hope this finds you well", short paragraphs. Better. Version three adds a persona paragraph and a few example phrases. The output now avoids the worst tells and still does not sound like you, so you rewrite the draft anyway, which is the thing the instructions were meant to end.

Two things are happening. First, the rules you wrote are mostly negative ("never say X"), and removing tells produces clean, anonymous text rather than your text. Second, the positive rules are adjectives, and adjectives have no numbers in them. "Concise" does not tell the model whether your sentences average eleven words or nineteen. "Warm" does not say whether you open with the person's name or with the topic. The model fills every unspecified dimension with its default, and its default is the average of everyone.

The Custom GPT route hits the same wall for the same reason. A "write like me" GPT built from a self-written system prompt is the same description in a different container, which is why the third failed GPT feels identical to the third failed instruction set. Our post on why your Custom GPT doesn't sound like you covers that variant in detail.

What Does a Style Profile Contain That Custom Instructions Cannot?

A Style Profile contains the countable features of your writing, stated as rules, with the evidence behind them. Sentence length, qualifier placement, transitions and punctuation are the four where readers recognise you, and none of them is an adjective. A profile states each one as a range or a rule the model can hit and you can check.

Sentence length. A range rather than an adjective: a typical sentence of 12 to 16 words, at least one under eight in every paragraph, none over 30. Those are example values; yours come from your samples. A model given a range can hit it. A model told to be "concise" picks its own.

Qualifier placement. Whether you hedge at all, and where. "I think we should ship Friday" and "We should ship Friday, I think" are the same opinion from two different people. Custom instructions almost never mention this. A profile measures it, because hedging position is one of the features that stays stable across everything a person writes.

Transitions. Whether you write "so" or "therefore" or start a new sentence with nothing. Whether a paragraph opens with the conclusion or builds to it. Whether you ever write "however". The model's defaults here are exactly the connectives that make a draft smell like AI, and a self-written instruction rarely lists them, because nobody notices their own transitions.

Punctuation. Colons per email, commas per sentence, parentheses or a separate sentence for asides, whether you use exclamation marks with clients. A profile counts these from your samples. You could not have typed them from memory.

Anti-patterns. The phrases that never appear in your writing, extracted rather than guessed. This is often the most identifying part of a profile, and the part self-written instructions get most wrong, because people ban the phrases they dislike in AI output rather than the phrases they themselves never use.

Context shifts. How the same person writes to a board and to a teammate. Custom instructions are one static block. A profile carries rules per audience, so the register moves when the recipient does.

Each of these comes from the same place: three to five pieces you wrote yourself, in the context you use most. Our post on how many writing samples AI needs explains why words per context is the unit that matters. The point for this comparison is simpler. A prompt describes how you hope to sound. A profile measured from your own writing is evidence of how you actually write, and evidence is the thing your custom instructions were missing.

Does a Better Model Make Custom Instructions Enough?

No, and the reason cuts the other way. Model capability and knowledge of your writing are separate layers. Labs improve the first with every release and cannot supply the second, because nothing about you is in the training data. A more capable model follows nuanced instructions more faithfully, which means it executes a measured profile better than the previous model did. It also executes a vague description more faithfully, generic wording included.

This is the part ChatGPT Plus users get backwards. The upgrade to a stronger model feels like it should fix voice, because the writing quality goes up. Quality and identity are different axes. Better models write cleaner, better-structured, more fluent average prose, and average prose is precisely the thing you keep rewriting. The better the model gets at following instructions, the more the output depends on what the instructions contain, so the gap between a description and a measurement widens with every release.

Our own measurements point the same way. When we deployed the same Style Profile across models in the cross-platform deployment guide, each model kept its own baseline personality, and the profile narrowed the spread between them from a 30 to 40 point gap down to 5 to 10 points on most dimensions. That is a measured instruction set doing its job on models that differ from each other. A description would have left each model's baseline in charge.

Custom Instructions vs Style Profile: Which One Do You Deploy?

Both, in layers. Custom instructions hold everything that is about the task, and the Style Profile holds everything that is about you. The profile's Runtime Block takes the voice slot in the field, and your own lines cover the rest.

Custom instructions remain the right place for:

• Facts about your situation. Role, company, audience, the products you write about. The "About you" box exists for this, and a profile never needs to know it.
• Task rules. Output format, length caps, "always give me three subject-line options", which language to answer in.
• Standing preferences. No preamble, no summary paragraph at the end, ask before assuming.

A Style Profile is the right place for anything that answers "how does this person write?", because that answer needs measurements and you cannot supply them by introspection. If you want to see how far self-description can be pushed before it plateaus, adjectives are not your voice walks through the four features that matter and how to count them yourself. Doing that by hand is a fair weekend project. MyWritingTwin.com does the same extraction across 50+ dimensions in minutes and hands back a profile you can paste.

Portability is the other reason to keep the profile separate from the field. Custom instructions live inside one ChatGPT account. The same Style Profile deploys as a Runtime Block in ChatGPT, as a full document in a Claude Project, as a Gem in Gemini, and as a system prompt in anything you build on an API. When the model that serves you best changes, and it will, your Writing Twin for ChatGPT, Claude and Gemini moves with you. Your old instructions stay behind in the account where you typed them.

How Do You Move From Custom Instructions to a Style Profile This Week?

Keep your task rules, replace the voice paragraph, and test on a real draft. The switch takes under an hour of your own time, most of it choosing samples.

1. Split your current instructions. Copy them into a text editor. Mark each line as task (format, length, language, facts about you) or voice (tone, style, phrases). Keep the task lines exactly as they are.
2. Choose three to five samples. Pieces you wrote yourself, untouched by AI, in the single context you rewrite most. For most ChatGPT Plus users that is client or team email. Skip the polished report you write twice a year.
3. Run the analysis. Upload the samples to MyWritingTwin.com. The Writing DNA analysis measures sentence length, qualifier placement, transitions, punctuation, anti-patterns and context rules, then produces the Master Prompt and the Runtime Block.
4. Paste the Runtime Block into the "How would you like ChatGPT to respond?" field, above your task lines. It fits inside the 8,000-character limit with room to spare.
5. Test on the draft you would normally rewrite. Ask for the email you were about to write anyway. Count what changed: sentence lengths, where the hedge sits, the opening line. If a dimension misses, adjust that one rule in the profile. You are now correcting a measurement, which is something you can do. Correcting an adjective was never possible.

The test at the end is the real difference between the two approaches. "Was that friendly enough?" has no answer, so custom instructions never converge. "Was the mean sentence length in range?" has one, so a profile does.

If you have rewritten your custom instructions three times and still edit every draft, the fourth rewrite will not fix it, because the input was never the problem. Build the profile from writing you already have. MyWritingTwin.com turns three to five samples into a Style Profile and a Runtime Block you can paste into ChatGPT, Claude or Gemini today, and the draft you would have rewritten tonight is the first test.