Project Instructions ChatGPT Actually Follows: A Voice-First Template

Q: What should I put in ChatGPT project instructions?

Put the things that stay constant across every chat in that project: who the reader is, what the output is for, the format you want, and how you write. The last part carries the most weight and is the part most people get wrong. Replace tone adjectives with measurements from your own writing: typical sentence length, where the ask sits, which words you never use, how you open and close.

Q: Why does ChatGPT ignore my project instructions?

Three usual causes. The instruction is vague enough that the model can satisfy it while doing something else, as with 'be concise'. Two instructions contradict each other and the model picks one. Or the instruction describes an effect rather than a behaviour, so there is nothing concrete to follow. Rewriting the rule as something countable fixes most of these.

Q: Is there a character limit on ChatGPT project instructions?

Yes. The project instructions field accepts several thousand characters, and OpenAI adjusts the exact figure over time, so check the field itself rather than trusting a number from an older article. The practical constraint is different: a measured voice block for one kind of writing fits in well under 1,500 characters, so the limit rarely matters once you stop writing paragraphs of adjectives.

Q: Can project instructions make ChatGPT sound like me?

They can get close if they contain the right kind of information. A paragraph describing your tone will not. A block of measured habits from your own writing, stated as rules the model can check itself against, will move the output much closer. The remaining gap is what a Style Profile built from your samples is designed to close, and it pastes into the same field.

You set up a ChatGPT Project, filled in the instructions field with care, uploaded a few files, and the first draft still reads like it came from the default assistant. The reason is almost never the model. The project instructions ChatGPT reads at the start of every chat in that project usually contain a job description for the assistant and almost nothing about the person whose name goes on the output.

Key takeaways: project instructions in ChatGPT

The instructions field is the only input a project reads on every single turn.Files are retrieved selectively. Memory is approximate. Instructions are always present.Most instruction blocks spend their characters on role and format and leave voice to adjectives."Professional yet friendly" selects a register the model learned from millions of writers and returns it to everyone who types it.A voice is a set of countable habits.Sentence length, where the ask sits, how you hedge, what you never say. Each one can be written as a rule the model can check.Knowledge files cannot substitute for those rules.The model sees your samples when it decides they are relevant and paraphrases what it sees. It does not measure them.A measured block fits in a fraction of the fieldand improves with every stronger model, because better instruction-following executes a precise spec more faithfully.

What Do Project Instructions in ChatGPT Actually Control?

Project instructions are a block of text that ChatGPT prepends to every conversation inside that project, in addition to whatever global custom instructions you have set. They are the one input the model reads in full, every turn, without deciding whether it is relevant. That makes the field the single most reliable place to put anything you need the model to do consistently.

OpenAI's own description of Projects in ChatGPT frames the feature as a workspace: chats, files and instructions grouped around one piece of ongoing work. The instructions field was designed with the workspace in mind, so the examples people copy tend to describe the work. "You are a marketing assistant for a B2B software company. Write in a professional yet approachable tone. Keep responses concise. Use British English." That is a job description for an employee who has never met you.

A project has three places where information about you can live, and they behave differently:

Input	When the model sees it	What it is good at	What it cannot do
Instructions	Every turn, in full	Rules, constraints, format, voice spec	Hold long reference material
Knowledge files	When retrieval judges them relevant	Facts, templates, past documents	Teach a pattern the model has to infer
Project memory	Approximate recall of earlier chats	Continuity on names and decisions	Anything that needs to be exact

The setup mechanics, including where each field lives and what the current limits are, are covered in the ChatGPT Projects setup guide. This post is about what goes in the field once you have found it.

The Project Instructions ChatGPT Follows Most Reliably

The instructions ChatGPT follows most reliably are the ones that can be checked. "Use British spelling" is checkable. "Never open with a greeting" is checkable. "Be warm" is not, so the model satisfies it in whatever way its training suggests, and that way is the same for every user who asked for warmth.

This is the gap between an instruction and a wish. An instruction tells the model what to do in a form where both of you could look at the output and agree whether it was done. A wish names an effect and leaves the method open. Modern models follow the first category with real precision and can do nothing with the second.

Sort a typical instruction block into those two piles and the pattern is stark. Role, audience, format, length and language are nearly always instructions. Voice is nearly always a wish. The part of the block that was supposed to make the output sound like you is the part written in the form the model cannot act on.

Where the Typical Instruction Block Leaks Your Voice

Here is an instruction block of the kind that circulates in templates, lightly anonymised from one we were sent for review:

You are my writing assistant for client emails. I run a small consulting practice. Write in a professional but friendly tone, confident without being arrogant. Keep emails concise and to the point. Avoid jargon. Always be polite. Use clear structure. Sign off with "Best regards".

Every sentence except the last two describes an effect. "Professional but friendly" is a register. "Confident without being arrogant" is a register. "Concise" has no number attached. "Avoid jargon" depends on whose jargon. "Polite" and "clear" are judged by the model, using its own defaults.

Now compare it with what the same consultant's last twenty emails actually show when you count:

Median sentence of 13 words, almost none over 25.
The ask appears in the first or second sentence, never in the last paragraph.
Hedges ("I think", "probably") appear before recommendations and never after a decision.
No "however", no "please find attached", no exclamation marks.
One colon per email at most, used to introduce a list.
Closes with a one-line next step and the time it will happen, then "Best".

None of that is in the instruction block, so none of it reaches the model. The output is professional, friendly and concise by the model's definition, and the consultant reads it and says it does not sound like them. Both are correct.

The complete guide to ChatGPT Projects covers layered templates for workflows. The leak described here is upstream of any template: it is about which kind of sentence you fill the template with.

Why Tone Adjectives Return the Same Draft for Everyone

A tone adjective points at a region of the model's training data, and the model answers with the centre of that region. "Professional yet friendly" is a label thousands of writers would accept for themselves, so the text that comes back is the average of thousands of writers. The model is not failing. It is doing exactly what the label allows.

The model does not know your tone. It knows words that describe tone, and it knows what text usually sits near those words. Your tone is the particular way you shorten a sentence when you are being direct, where you put the qualifier when you are unsure, how often you use a colon, whether your paragraphs run two sentences or five. Those are habits with numbers attached, and the model never saw the numbers because you gave it an adjective instead.

A quick way to feel the difference: ask ChatGPT, in a fresh chat with no project, to write a short client email "in a professional yet friendly tone", and ask a colleague to do the same in their account. The two drafts will share sentence rhythm, hedging pattern and closing shape, because both of you sent the same pointer to the same average. We covered this failure in why adjectives are not your voice: a voice lives in sentence length, qualifier placement, transitions and punctuation, and a label carries none of those.

This is also why upgrading to a newer model does not fix a project that sounds generic. Model capability and knowledge of you are separate layers. The lab improves the first and has never seen your writing. A stronger model follows "professional yet friendly" more faithfully, which means a more polished version of the same average. Feed that same stronger model a precise spec and it executes the spec more faithfully too, which is the direction you want the improvement to run.

If you want to know how far your current project is from your real writing before changing anything, the Humanity Score test compares an AI draft against your own text in about a minute and tells you which habits the draft missed.

Why Knowledge Files Do Not Fix the Instructions

Uploading ten of your best emails as knowledge files feels like it should solve the problem, and it helps less than it should. Files in a project are retrieved, not studied. When you ask for a new email, the model may pull a passage from one sample if retrieval judges it relevant, then paraphrase what it pulled. It does not read all ten, count the sentence lengths, notice that you never use "however" and derive a rule.

That inference step is the one the model skips. Ten samples contain your habits and everything particular to those ten situations, and the model cannot tell which patterns are you and which were a Tuesday. A measured summary resolves that, and the summary belongs in the instructions field, where it is read every turn.

The right division of labour is simple. Files hold material the model should quote or reference: a pricing sheet, a past proposal, the house glossary. Instructions hold the rules for how the output is written, stated so the model can check them. Our earlier piece on knowledge files and personas in ChatGPT Projects goes through what files do well and where their voice replication ceiling sits.

How Should You Structure Project Instructions for Voice?

Structure the block in four short sections, in this order: reader, job, format, voice. Reader and job set the context the model needs to make choices. Format fixes the shell of the output. Voice is where the countable rules go, and it should be the longest section even though most templates make it the shortest.

Putting voice last also keeps you honest. When the section is a list of checkable rules, you can see at a glance whether any of them is really an adjective in disguise.

Each voice rule should pass a one-question test: could you look at a draft and say whether the rule was followed? "Typical sentence 12 to 14 words" passes. "Natural rhythm" fails. "Ask in the first two sentences" passes. "Get to the point" fails. The rules you keep will look dry and mechanical on the page. That is the sign they are doing their job.

A Voice-First Project Instructions Template

Below is a template built on that structure, with the consultant's measurements filled in. Replace the numbers with your own; the shape stays the same. It runs to about 1,100 characters, which leaves most of the field free.

READER
Existing clients and warm prospects at mid-sized companies. They know my work. They are busy.

JOB
Draft client emails from my brief. One email per request. Do not add content I did not ask for.

FORMAT
Plain text, no headings, no bullet points unless I ask. 60 to 140 words. Subject line first.

VOICE (follow these exactly)
- Typical sentence 12 to 14 words. Never over 25.
- State the ask in the first or second sentence.
- Hedge only before a recommendation ("I think we should"). Never hedge after a decision.
- Never use: however, please find attached, I hope this finds you well, exclamation marks.
- At most one colon per email, and only to introduce a list.
- Close with one line naming the next step and when it happens, then "Best" and my first name.

Three things to notice. There is not a single adjective in the voice section. Every rule names a quantity, a position or a word. And the rules are specific enough that a draft either follows them or does not, so when a draft drifts you can point at the rule it broke rather than saying it felt off.

The measurements come from writing you have already done. Pull up your last fifteen to twenty pieces of the kind this project will produce and count six things: words per sentence, where the ask sits, how you hedge, words you never use, punctuation habits, and how you open and close. An hour of counting gets you a block like the one above. A Style Profile built from your samples does the same measurement across many more dimensions and produces a Writing Twin for ChatGPT, Claude and Gemini that pastes into this same field; the comparison in custom instructions vs Style Profile covers where a hand-counted block stops and a measured profile continues.

Three Rules for Keeping Instructions Inside the Limit

The field is finite and people hit the ceiling when they paste a long persona paragraph followed by sample emails. Three rules keep the block short without losing anything the model uses.

Samples go in files, rules go in instructions. A sample in the instructions field costs hundreds of characters and teaches the model nothing it can act on directly. A rule derived from that sample costs a line. If you want the model to be able to quote a past email, upload it as a file.

One project per kind of writing. Client emails, proposals and LinkedIn posts have different sentence lengths, different hedging and different closes. One project that tries to cover all three ends up with a voice section full of conditionals, and the model resolves conflicts by picking one. Three projects with three short voice sections behave better and each stays well inside the limit.

Cut every rule you cannot check. If a line in the voice section is an adjective, it is costing characters and contributing an average. Delete it or convert it. "Be direct" becomes "ask in the first two sentences". "Concise" becomes a word range. "Warm" usually turns out to mean one specific opening habit you can name.

A Five-Minute Test for Your Current Project

Take an email you actually sent last week, the kind this project is meant to produce. Write a two-line brief for it: who it went to, what you needed, what they already knew. Open the project, paste the brief, and ask for the email.

Put the draft next to the one you sent and count:

Words per sentence, shortest and longest
Where the ask appears
Number of hedges and where they sit
Any words in the draft you would never write
Punctuation that differs: colons, dashes, exclamation marks
Opening line and closing line

If the numbers differ, the instructions did not carry your voice, however pleased the model sounds when you ask it directly. Rewrite the voice section as rules, run the same brief again, and count again. Two passes is usually enough to see the numbers converge on one email. Then try three more briefs in different registers, because a tone adjective often survives the first test and fails the second, and a measured block survives both.

The test replaces a feeling with a count. "Does it sound like me?" gets a different answer on different days. "Did it hit the sentence range and put the ask first?" has one answer, and when that answer is no, it names the rule to fix.

Frequently Asked Questions

What should I put in ChatGPT project instructions?

Put what stays constant across every chat in the project: who the reader is, what the output is for, the format, and how you write. The voice section carries the most weight and is the one most templates get wrong. Replace tone adjectives with measurements from your own writing, such as typical sentence length, where the ask sits, words you never use, and how you open and close. Keep each rule checkable.

Do project instructions override ChatGPT custom instructions?

Inside a project, the project instructions take precedence where the two conflict, and your global custom instructions still apply for anything the project does not address. The practical pattern is to keep global custom instructions short and general, then put the specific voice rules for each kind of writing in the project that produces it, so the rules closest to the output are the most precise ones.

Why does ChatGPT ignore my project instructions?

Usually one of three things. The instruction is vague enough that the model can satisfy it while doing something you did not want, as with "be concise". Two instructions contradict each other and the model picks one without telling you. Or the instruction names an effect, like "sound confident", rather than a behaviour, so there is nothing concrete to follow. Rewriting the rule as something countable fixes most cases.

Is there a character limit on ChatGPT project instructions?

Yes. The field accepts several thousand characters and OpenAI adjusts the exact figure over time, so check the field itself rather than trusting a number from an older post. In practice the limit rarely binds once you stop pasting samples and adjectives: a measured voice block for one kind of writing fits comfortably under 1,500 characters and leaves most of the field free for format and job rules.

Can project instructions make ChatGPT sound like me?

They can get close when they contain the right kind of information. A paragraph describing your tone will not do it, because the model maps adjectives to an average. A block of measured habits from your own writing, stated as rules the model can check itself against, moves the output a long way. The gap that remains is what a Style Profile built from your samples is designed to close, and it pastes into the same field.

Find out how far your current project is from your real writing with the Humanity Score test, or build a Style Profile from your own samples and paste it into the instructions field of the ChatGPT Project you already use.