Instructions Personnalisées ChatGPT: Make Them Sound Like You

TL;DR: If you searched for instructions personnalisées ChatGPT, the setting is under your profile icon, labelled Personnaliser ChatGPT, with two fields: one about you, one about how ChatGPT should respond. Most people fill the second field with a tone label ("professionnel mais chaleureux") and get the same generic French as everyone else who typed it. AI does not know your tone; it knows words that describe tone. Your voice lives in sentence length, qualifier placement, transitions and punctuation, and it shifts between French and English. Give ChatGPT those measurements, per language, and the fields finally do their job.

You found the setting, you wrote "réponds en français, ton professionnel mais chaleureux, sois concis", and ChatGPT still opens every email with "J'espère que ce message vous trouve bien." The setting works. The instruction does not.

This guide covers where the instructions personnalisées live, what each field is for, and why the words most people put in them cannot produce their voice. Then it shows what to write instead, in French and English.

Instructions Personnalisées ChatGPT: Where the Setting Lives

The setting sits behind your profile icon at the bottom left of the ChatGPT window. Click it, choose Personnaliser ChatGPT (Customize ChatGPT in the English interface), and a panel opens with the two text fields that make up the instructions personnalisées. Save, and they apply to every new conversation from then on.

A few details people miss on the first visit:

• The instructions apply to new chats only. Conversations opened before you saved keep running without them.
• There is a toggle to enable them for new chats. If your instructions seem ignored everywhere, check that it is on.
• On mobile, the path is the same: profile icon, then Personnaliser ChatGPT.
• The fields sync across devices, so a change on your laptop reaches your phone.
• Custom GPTs and Projects carry their own instructions, which sit alongside your personal ones rather than replacing them.

OpenAI documents the feature in its help center article on custom instructions. The article explains the mechanism. It says nothing about what to write, which is where every French-speaking professional we have talked to gets stuck.

The Two Fields and Their Limits

The first field asks what ChatGPT should know about you. The second asks how ChatGPT should respond. They have different jobs and different sizes.

The size difference matters. The style field holds more than five times the context field, which tells you where OpenAI expects the work to happen. Most users invert this. They write three paragraphs about their job in the first field and one line of adjectives in the second.

Writing in French costs more characters than writing in English for the same instruction, roughly a fifth more in our own drafts, so a bilingual set of instructions needs to be edited tightly. The 1,500-character field fills fast if you describe both your French and your English context in prose. Lists survive the cut better than sentences.

A Tone Label Is Not a Voice

"Professionnel mais chaleureux" describes thousands of people. So does "clair et direct", "bienveillant", "synthétique". These are the phrases that fill the second field in almost every set of instructions personnalisées we have reviewed, in French and in English, and they all produce the same text.

The reason is structural. A language model has no access to your tone. It has access to the words you used to describe your tone, and it maps those words to the register that appeared most often near them in training. "Professionnel mais chaleureux" points to one learned average of French business writing. Everyone who types the phrase gets the same average. We covered the general mechanism in tone adjectives are not your voice; the French case is sharper because the French corpus of business writing is smaller, so the average is narrower.

Try this on your own drafts. Take two colleagues who would both describe themselves as "professionnel mais chaleureux". Put their emails side by side. One opens with the request and closes with a single line of warmth. The other spends a paragraph on context before asking anything. One writes sentences of eight to twelve words; the other regularly passes thirty. One never uses an exclamation mark; the other uses one per message. Same label. Different voices.

The label survives because it feels like a description. It is a category. A voice is a set of measurements, and categories do not carry measurements.

You can see how far your current instructions are from your real writing in about a minute with the Humanity Score test. Paste a draft ChatGPT wrote under your instructions, paste one you wrote yourself, and read the gap.

Why Does "Professionnel mais Chaleureux" Produce the Same Text for Everyone?

Because the model resolves the phrase to a register, and a register is a shared default, not a personal one. Your instructions become a lookup, and the lookup returns the same entry for every user who typed the same key.

Consider what the model actually does with the phrase. It raises the probability of certain openings ("J'espère que vous allez bien"), certain closers ("N'hésitez pas à revenir vers moi"), a moderate sentence length, a moderate density of politeness markers, and the conditional mood in requests ("Serait-il possible de..."). Those choices are the statistical centre of French professional email. They are not wrong. They are simply nobody's.

Three things you might think rescue the label do not:

1. Adding more adjectives. "Professionnel, chaleureux, direct, humain, précis" narrows nothing. Each word pulls toward its own average, and the result is a blend of averages.
2. Adding "comme moi" or "dans mon style". The model has no sample of your style in the fields, so the instruction refers to nothing it can see.
3. Correcting it in the chat. Corrections shape the rest of that conversation and die with it. The next chat reloads the label.

The fix is to replace the category with the measurements the category was standing in for.

The Four Things ChatGPT Can Actually Execute

A voice lives in sentence length, qualifier placement, transitions and punctuation. Those four are the dimensions a model can follow as instructions, because each one is countable in the output.

Sentence length. State a range and a typical value. "Phrases de 10 à 20 mots, moyenne autour de 14. Une phrase de moins de 6 mots par paragraphe, au maximum." That instruction is checkable. "Sois concis" is not.

Qualifier placement. Where your hedges sit, and how many. Some writers open with the hedge ("Je pense que nous devrions..."), others close with it ("...ce qui reste à confirmer"). Some never hedge a decision and hedge every forecast. Tell the model which you do.

Transitions. The connectors you use and the ones you never use. If you never write "par ailleurs" or "en outre", say so, because the model reaches for both under any professional label. List the two or three you actually use ("d'où", "concrètement", "donc" in second position).

Punctuation. Exclamation marks, semicolons, parentheses, the colon before a list, whether you write "..." or a full stop. In French, add the spacing rule you follow before "?" and ":" and whether you use « » or straight quotes. The model defaults to the typographic norm, which may not be yours.

Add vocabulary on top of the four: five phrases you reach for, five you never write. Keep the never-list short. Long ban lists get ignored past the first dozen entries.

Register Changes Between French and English, and the Model Flattens It

Your French tutoie or vouvoie. Your English does neither, so it carries formality through word choice and sentence shape instead. A single tone label in the instructions personnalisées treats both languages the same and misses the shift in each.

This is the part of the problem that only shows up for people who write in more than one language, and it is the part no tone label can address. We measured the flattening effect in how AI writes differently across languages. Under one generic instruction, the model produced French that was more formal than the writer's own and English that was more casual than the writer's own, pulling both toward the middle of its training distribution.

Some specifics the model needs per language:

Notice that the English sentence length target is shorter. Most bilingual writers we have profiled write shorter sentences in their second language, and the model will not know that unless told. One label, one length. Two languages, two sets of numbers.

How Do You Write Custom Instructions That Carry Your Voice in French and English?

Measure ten pieces of your own writing per language, state the findings as constraints, and put the constraints in the second field under a heading for each language. The whole set fits in well under 8,000 characters.

The method, step by step:

1. Collect ten samples per language that you wrote and were happy with, in the register you need most. Emails count. Slack messages count if that is where you write.
2. Count words per sentence across all ten. Note the shortest, the longest and the typical value.
3. Mark every hedge and where it sits in the clause. Count them per paragraph.
4. List the transitions you used and the professional ones you never used.
5. Note your punctuation habits: exclamation marks per message, semicolons, parentheses, quote style, list style.
6. Write each finding as an instruction, not an adjective, under a "Français" heading and an "English" heading in the second field.
7. Put context in the first field: your role, who you write to, which language goes to whom. Keep it to bullets.
8. Test with an email you already sent. Give ChatGPT the brief you would have given a colleague, then count the same dimensions in its draft and yours.

Step 8 is the one people skip and the one that pays. If the numbers match on one email in each language, try two more in different registers. If they drift, tighten the instruction that drifted, not the whole set.

A Worked Example: Before and After

The before version is the second field as most French-speaking professionals write it. The after version is the same writer, measured.

Before:

Réponds en français sauf si je t'écris en anglais. Ton professionnel mais chaleureux. Sois concis et clair. Évite le jargon.

After:

Français. Vouvoiement par défaut. Commence par l'objet du message, sans formule d'ouverture. Phrases de 10 à 20 mots, une phrase courte (moins de 6 mots) par paragraphe au maximum. Pas de conditionnel pour une décision prise; conditionnel pour une demande à un supérieur. Transitions autorisées: "donc", "concrètement", "d'où". Jamais: "par ailleurs", "en outre", "n'hésitez pas". Une seule formule de clôture: "Bien à vous". Pas de point d'exclamation. Guillemets « ».

English. Open with the ask. Sentences of 8 to 16 words. One hedge per paragraph at most, placed at the end of the clause ("...which we still need to confirm"). Never "I hope this finds you well", never "please don't hesitate". Sign off "Best," on its own line. No exclamation marks. Paragraphs of two to four sentences.

The after version is longer and still under 900 characters. Every line in it can be checked against a draft. Nothing in the before version can.

Run both through the same brief ("relance au client pour la facture de juillet, il a déjà reçu deux rappels, garder la relation") and the difference is visible in the first line. The before version opens with "J'espère que vous allez bien". The after version opens with the invoice.

Better Models Execute Your Instructions Better, Not Guess Better

Model capability and user-data conditioning are separate layers. Each new ChatGPT model follows nuanced instructions more faithfully than the last, and none of them adds a sentence of your writing to the model. The better the model gets at following precise instructions, the more a precise set of instructions personnalisées returns.

This cuts against a hope we hear often from French-speaking users: that the next model will "understand" the tone label properly. It will understand the label exactly as well as the current one does, because the label has no more information in it than it had before. What improves is execution. Give a stronger model "phrases de 10 à 20 mots, pas de conditionnel pour une décision" and it hits those constraints more reliably than a weaker one. Give it "professionnel mais chaleureux" and it produces a slightly better-written version of the same average.

We laid out the two-layer argument in full in why better AI models still don't know your voice. The practical consequence for your instructions personnalisées is simple: every character you spend on a measurable constraint appreciates with each release. Every character you spend on an adjective does not.

Custom Instructions Versus a Style Profile

Custom instructions are the container. A Style Profile is the content. The container is free and lives in ChatGPT's settings; the content decides whether the output sounds like you, and you can build it by hand with the method above or extract it from your writing.

The manual method works. It takes an afternoon per language, it produces rough numbers, and the numbers are yours. Its weakness is coverage: ten samples show one mood, and your voice has several. The register you use with a client in Lyon differs from the one you use with a colleague in London, and both differ from a board memo.

A Style Profile from MyWritingTwin.com does the measurement across a larger sample set, per language, and states the result as constraints ready to paste. The Writing DNA analysis covers the same four dimensions (sentence length, qualifier placement, transitions, punctuation), plus vocabulary and register shifts between your languages. The output is a block of text. It goes into the second field of your instructions personnalisées, into a Claude Project, or into Gemini's saved info without changes. One profile, every model, both languages. We compared the two approaches in detail in custom instructions versus a Style Profile.

Whichever route you take, the principle is the same. Stop describing your voice and start measuring it. The fields have been waiting for numbers all along.

Frequently Asked Questions

Where are the instructions personnalisées in ChatGPT?

Click your profile icon at the bottom left of ChatGPT, then choose Personnaliser ChatGPT (Customize ChatGPT in the English interface). Two text fields open: one about you, one about how ChatGPT should respond. They apply to every new conversation once saved, and you can switch them off per chat. The setting is available on free and paid plans, on web and mobile, and the fields sync across your devices.

Do custom instructions work in French?

Yes. You can write both fields in French, and ChatGPT follows them in whichever language you write your prompt. The limit is the content, not the language. An instruction like "ton professionnel mais chaleureux" maps to one generic French register that is identical for everyone who types it. Instructions that state sentence length, register per language and punctuation habits produce text that sounds like you in French and in English.

Why does ChatGPT ignore my custom instructions after a few messages?

Long conversations push the instructions further from the text the model is generating, and vague instructions drift first. "Be concise" has no measurable target, so the model relaxes toward its default length. Instructions phrased as measurable constraints (sentence length range, one hedge per paragraph, no exclamation marks) hold longer because each one can be checked against the draft. When drift starts, open a new chat: the instructions reload at full strength.

Is a Style Profile different from custom instructions?

A Style Profile is the content you paste into the custom instructions fields, built from measurements of your own writing rather than from adjectives. Custom instructions are the container ChatGPT provides. MyWritingTwin.com extracts a Writing DNA analysis from samples you wrote, states it as constraints the model can execute, and the same block works in ChatGPT, Claude and Gemini. The container is free. The profile is what decides whether the output sounds like you.

Measure your voice once, in each language you write, and paste the result where ChatGPT can execute it. If you want to know how far your current instructions are from your real writing before you start, the Humanity Score test takes about a minute.