Custom GPT Doesn't Sound Like Me: Why, and the Fix That Works

TL;DR: Your custom GPT doesn't sound like you because nothing about you ever reached it. A Custom GPT takes four inputs (instructions, knowledge files, conversation starters and a model choice) and none of them carries your voice unless you put it there in a form the model can act on. Tone adjectives point at a generic register. Knowledge files are retrieved, not absorbed. A newer model executes vague instructions more faithfully, which means a cleaner version of the same average. AI does not know your tone; it knows words that describe tone. The fix is to replace the description in the instructions field with measurements from your own writing: sentence length, qualifier placement, transitions and punctuation.

You spent an evening building a Custom GPT for client emails. You named it, wrote a paragraph about your style, uploaded five emails you were proud of, and picked the best model. The first draft it produced could have come from any assistant on any laptop in your building.

The GPT did what you asked. What you asked did not contain you.

Why Doesn't My Custom GPT Sound Like Me?

Because the GPT received a description of the effect you want and a few examples of the effect, and it has to invent the mechanics on its own. It invents them the same way for everyone, from the average of everything it was trained on. Your voice, the part a reader would recognise, was never handed over.

Walk through what most people put into the builder. The instructions field gets a paragraph like "You write like Priya: warm but direct, concise, confident, no corporate filler." The knowledge section gets a handful of past emails. The model dropdown gets whatever is newest. Each of those steps feels like it is teaching the GPT who you are. Each one leaks your voice on the way in.

There are four leaks, and they stack:

1. The instructions field holds adjectives. An adjective is a pointer to a register the model already knows. "Concise" points to the model's own idea of short.
2. Knowledge files are retrieved, not absorbed. The model looks them up when it decides they matter and paraphrases what it finds.
3. The model layer is generic by design. A new model release makes every Custom GPT better at following instructions and adds nothing about you.
4. Long chats drift. Your instructions sit at the top of the context, the model's own output accumulates below, and by the tenth turn the default has reclaimed the draft.

The rest of this post takes each leak in turn, then shows what to put in the instructions field so the GPT has something of yours to work with.

What Does a Custom GPT Actually Receive From You?

Four inputs, and only one of them can carry your voice. OpenAI's own guide to creating a GPT lists them: instructions, knowledge, conversation starters, and capabilities such as browsing or code. The instructions field is where behaviour is set. Everything else is either reference material or a feature toggle.

Most people spread their effort across all four and put the least specific content into the one that matters. The instructions field gets a tone paragraph. The knowledge section gets the real evidence, where the model can only reach it by lookup.

Look at what "warm but direct, concise, confident" gives the model to work with. During training it saw enormous amounts of text described with those words: HR announcements, customer-support replies, LinkedIn updates, sales follow-ups. The phrase now points to the statistical centre of all of it. Type it and you get the centre. A colleague who types the same phrase gets the same centre. The two of you write nothing alike, and the GPT has no way to know that, because the label was the only thing either of you provided. We took that argument apart in detail in why tone adjectives are not your voice. The Custom GPT version of the problem is the same problem with a nicer icon.

Why Don't Knowledge Files Fix It?

Because knowledge is a lookup, and a voice is a set of constraints that has to apply to every sentence. Uploading your emails gives the GPT a library to consult. It does not change how the GPT writes when it is not consulting the library, which is most of the time.

Consider the mechanics. When you ask your GPT for a follow-up email, the model decides whether any uploaded file is relevant to the request. If it decides yes, it pulls the relevant passages into context and uses them as reference. If it decides no, the files never enter the conversation. Either way, the files inform the content of the answer. They do not set the rhythm of the prose.

Even when the files are pulled in, the model faces an inference problem it cannot solve from five samples:

• Habit or mood? One email opens with a personal line. Was that your pattern or that one Tuesday? Five samples cannot separate the two.
• Which register? A client apology and an internal status update in the same folder pull the model in opposite directions, and it averages them.
• What to reproduce? The model sees short sentences and an early ask, and it may reproduce those. It will not notice that you never use "however", because absence is invisible in a sample.

There is a ceiling on the instructions field too. Paste three full emails into it and the space for actual rules is gone. Move them to knowledge, and you are back to lookup. Our guide to using a Custom GPT inside a ChatGPT Project covers the container question in depth: which of instructions, Projects, memory and GPTs should hold what. The short version is that samples are evidence, and the model still has to draw the conclusion. Draw it yourself and hand over the conclusion.

Will a Better Model Fix My Custom GPT?

No, and the reason is structural rather than a matter of waiting for the next release. Model capability and user-data conditioning are separate layers. The lab improves the first with every version, and it cannot supply the second, because the second is made of your writing and the lab has never seen it.

A more capable model follows nuanced instructions more faithfully. That is the whole of what an upgrade delivers to a Custom GPT. Feed it "warm but direct" and it produces a more polished, more consistent, more fluent rendering of the generic warm-but-direct register. The draft reads better and sounds no more like you than it did on the old model. Several readers told us their GPT got worse after a model switch, and the likely cause is the same: the new model executes the adjectives more thoroughly, so the average is even more average.

The same property works in your favour once the instructions carry something precise. A profile that says "sentences average 13 words, range 5 to 26, ask in the first line, hedge only before recommendations, never 'however'" is a set of instructions a stronger model can hit more accurately than a weaker one. The better the model gets at following nuanced instructions, the better it executes your profile. We laid out the evidence for this in why better AI models still don't know your voice: each generation narrows the gap between instruction and execution, and none of them narrows the gap between the model and you. Only your data does that.

So the model dropdown is the last thing to change, not the first. Fix what the model is being asked to execute, then let every future release execute it better.

What Should Go in the Instructions Field Instead?

Measurements from your own writing, stated as rules the model can follow without interpreting. A voice lives in sentence length, qualifier placement, transitions and punctuation, and every one of those can be counted from emails you have already sent.

Here is what each one looks like as an instruction rather than an adjective:

Sentence length. Not "concise" but a distribution. "Typical sentence 12 to 14 words. Drop to 4 to 6 words when stating a decision. Allow one sentence over 25 words per paragraph when setting context." A model given "concise" writes uniform fourteen-word sentences forever. A model given the range switches gears the way you do.

Qualifier placement. "Hedge before a recommendation ('I'd suggest', 'probably worth'). Never hedge after a decision has been stated. Do not use 'I think' in the first line." Two writers with identical vocabularies differ here on every paragraph, and readers feel the difference without naming it.

Transitions. "Start sentences with 'So' and 'And' when the previous sentence sets them up. Never use 'however', 'moreover' or 'that said'. Open paragraphs with the conclusion." The model's default transitions are the ones it saw most, and they are precisely the ones that make a draft smell like AI.

Punctuation. "At most one colon per email. Asides go in parentheses, never in a separate sentence. No exclamation marks in client email. No semicolons." Punctuation habits are the most stable dimension in most people's writing and the one tone adjectives cannot express at all.

On top of those four sit vocabulary and avoidance: the words you reach for and the phrases you would never write. Both are lists, and a list is an instruction.

The manual version takes about an hour:

1. Collect ten to fifteen pieces you wrote and were happy with, in the register the GPT is for.
2. Count words per sentence across all of them. Note the typical value, the shortest, the longest.
3. List every hedge and where it sat in the clause.
4. List the transitions you used and, separately, the ones you never used.
5. Note colon, comma, parenthesis and exclamation habits.
6. Write the findings as instructions with numbers in them. Delete every adjective.

That block goes into the Custom GPT instructions field and nowhere else. Keep the knowledge section for things that are actually reference material: product names, pricing, the client's history. Our complete guide to Custom GPT instructions has the field-by-field layout, including where a profile like this sits relative to task instructions and output format.

This is also what a Style Profile from MyWritingTwin.com is: the measurement step done for you, across more dimensions than the four above, from samples you upload once. The output is a Writing DNA analysis expressed as constraints and packaged as a block you paste into the instructions field. Your Writing Twin for ChatGPT, Claude and Gemini is that block, deployed inside the tools you already pay for, and the same block that fixes a Custom GPT also fixes a Claude Project or a Gemini Gem. Our explanation of how style extraction works walks through the full list of what gets measured.

Two properties of a measured profile matter here. It can be checked: "was that direct enough?" has no answer, and "was the mean sentence length in range?" does, so when a draft drifts you count, find the miss, and correct one rule. And it outlives the model: the measurements describe you, so they hold when the model changes, and each stronger model executes them more faithfully than the last.

How Do You Test Whether Your Custom GPT Knows You?

Give it the brief for an email you already sent and compare by counting. Five minutes, no tools, and the result is a number rather than a feeling.

Pick an email from last week that went well. Write a two-line brief: who it was to, what you needed, what they already knew. Paste the brief into your Custom GPT and ask for the email. Put its draft beside yours and count:

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

If the numbers differ, the GPT does not know you, whatever it tells you when you ask it "do you understand my style?" (it will say yes; it always says yes). If they match on one email, run three more in different registers: a client apology, an internal update, a cold introduction. Adjective-based GPTs usually pass the first test and fail the second, because the paragraph was written for one mood and you have several.

Run the same test after you replace the paragraph with measurements. The numbers converge, and they keep converging as you tighten individual rules. That convergence is the entire difference between a GPT that has your name on it and a GPT that has you in it.

Your Custom GPT was never broken. It was empty. Fill the instructions field with what your writing actually does, and the container starts doing its job.

See how far your current GPT is from your real writing with the Humanity Score test, or build a Style Profile from your own samples and paste it into the GPT you already made.