Best ChatGPT Practices: 7 Rules That Make It Write Like You

Most lists of best ChatGPT practices tell you how to phrase a prompt. Almost none of them deal with the thing that decides whether the output sounds like you: what ChatGPT knows about your writing before you type anything.

That gap explains a familiar experience. You follow every prompting tip, the answer is accurate and well organised, and you still rewrite half of it because it reads like a press release. The fix is not a cleverer prompt. It is a setup.

TL;DR: the best ChatGPT practices in one list

Treat model capability and your own data as two separate layers.OpenAI improves the first with every release. Only you can supply the second.Stop describing your tone with adjectives."Professional yet friendly" matches thousands of writers. Give ChatGPT measurable rules: sentence length, qualifier placement, how you open and close.Put standing rules in custom instructions, subject context in Projects, and voice evidence in both.Each layer has one job.Edit the output once, then fix the instructionso the same correction never happens twice.Retest after every model update.A better model follows a precise profile more faithfully, and follows a vague one more confidently in the wrong direction.The fastest way to get the measurable rules is to extract them from your own writing with aStyle Profilebuilt for ChatGPT, Claude and Gemini.

Why Most ChatGPT Practices Advice Stops at the Prompt

Prompting advice is easy to demonstrate and easy to share. You show a weak prompt, a strong prompt, and the difference in output. It makes a good screenshot.

The problem is what a prompt can and cannot carry. A prompt describes the task: summarise this, draft that, make it shorter. It can also describe how you hope to sound. What it cannot do is show ChatGPT how you actually write, because you are not inside the prompt. Your writing is.

So the usual advice optimises the one layer that was already working. ChatGPT has been good at following task instructions for a long time. Where it struggles is reproducing a specific person, and no amount of "write in a confident, warm, concise voice" closes that distance. The seven practices below start from that distinction.

Practice 1: Separate Model Capability From Your Own Data

The model and your data are different layers, and only one of them improves on its own. Every new release makes ChatGPT better at following nuanced instructions. None of them teaches it how you write a difficult email, because that information lives in your sent folder, not in the training data.

This matters because people draw the wrong conclusion from model upgrades. The reasoning goes: the model got smarter, so I need less setup. The opposite is true. A model that follows instructions more faithfully executes a precise profile better than its predecessor did. It also executes a vague profile more confidently, which means a more polished version of the wrong voice.

We wrote about this pattern in detail in why better AI models still don't know your voice. The short version: capability is the lab's job, conditioning on your data is yours, and the second job gets more valuable as the first one gets done.

Practical rule: when a new model ships, do not simplify your setup. Rerun your test set (Practice 7) and expect the gap between a precise profile and a vague one to widen.

Practice 2: Replace Tone Adjectives With Measurable Rules

"Professional yet friendly" is a description thousands of people would give of their own writing. It cannot identify any one of them. ChatGPT does not know your tone. It knows the words that describe tone, and it maps those words to the average of everyone who has used them.

A voice lives somewhere else. It lives in average sentence length and how much that length varies. In where you put qualifiers ("probably" at the start of a clause or buried in the middle). In how you move between paragraphs: with a connective, with a question, or with nothing. In whether you use semicolons, parentheses, or neither. These are countable, and ChatGPT follows countable rules far better than it follows moods.

Compare the two versions of the same instruction:

Adjective version	Measurable version
Write in a confident, warm, professional tone.	Average 14 words per sentence, with at least one sentence under 7 words in every paragraph.
Be concise.	No paragraph over 4 sentences. Cut every opener that restates the subject line.
Sound natural and approachable.	Address the reader as "you" and never use "one" as a pronoun. Contractions are fine except in the closing line.
Avoid being too formal.	Open with the point. Never open with "I hope this finds you well" or any variant.

The right column is what a Style Profile contains. The left column is what most people type into custom instructions and then blame ChatGPT for ignoring. We go further into this in tone adjectives are not your voice.

Practice 3: Put Standing Rules in Custom Instructions, Not in Every Prompt

Anything that is true of all your writing belongs in custom instructions, where it applies to every conversation. OpenAI's own custom instructions documentation describes the two fields: one about you, one about how you want ChatGPT to respond. The second field is where your measurable rules go.

Repeating rules in each prompt fails for two reasons. You forget some of them on the days you are busiest, which are the days the output matters most. And ChatGPT weighs a long, repeated preamble less consistently than a stable instruction set it sees at the start of every session.

Structure the field in three blocks: ALWAYS, NEVER, and CONTEXT SHIFTS. ALWAYS holds your baseline rules. NEVER holds the ChatGPT defaults you refuse to accept, named exactly ("Please don't hesitate to reach out" works better than "avoid clichés"). CONTEXT SHIFTS states how the rules move when the audience changes. We publish six filled-in versions of that structure in our ChatGPT custom instructions templates.

Practice 4: Use Projects to Hold Context, Not to Hold Your Voice

Projects are for material that belongs to one body of work: a client, a product, a book. Put the background documents, the glossary, the audience notes and the format examples there. Do not put your voice there and nowhere else.

The reason is portability. Your voice is the one thing that stays constant across every Project, every chat, and every other AI tool you use. If it lives inside a single Project, it does not follow you to the quick question you ask in a plain chat, and it does not follow you to Claude or Gemini at all. Custom instructions carry the voice. Projects carry the subject. Our complete guide to ChatGPT Projects walks through the split in detail.

One useful exception: a Project can override voice rules for a register that is genuinely different. Your board memo voice and your Slack voice may need separate rule sets. The override lives in the Project. The baseline still lives in custom instructions.

Practice 5: Give Examples of Your Writing, Then Say What to Notice

Examples beat descriptions, but only when you tell ChatGPT what the example proves. Paste three paragraphs you wrote and say "write like this" and you get surface imitation: a similar topic, a similar length, a generic voice underneath.

Paste the same three paragraphs with a note ("notice the sentence that opens each paragraph is under 8 words; notice I never use a colon before a list") and the model has something it can apply to new material. The annotation turns an example into a rule.

This is also the answer to a common objection: "I don't know what my own patterns are." Most people do not. They are unconscious by definition. You can find them by hand with a word processor and an hour, or you can run your samples through a tool that measures them. Either way, the output you want is the right-hand column from Practice 2, not a pile of unlabelled paragraphs.

Practice 6: Edit the Output Once, Then Fix the Instruction

Every correction you make to ChatGPT output is a rule you have not written down yet. If you change "at your earliest convenience" to "by Friday" for the third time this week, the problem is not the draft. It is a missing line in your NEVER list.

Keep a running file for one week. Each time you edit a draft, write down the edit in rule form. At the end of the week, move the rules that appeared more than once into custom instructions. Most people find that a short list of rules covers the large majority of their edits, and that the output needs noticeably less rework in week two.

The discipline here is to stop treating edits as the cost of using AI and start treating them as data about your setup. We make the same argument, from the other direction, in stop editing AI output.

Practice 7: Retest Your Setup After Every Model Update

A model update changes how your instructions are interpreted, even when the instructions have not changed. Keep a fixed test set and run it after every release.

A good test set is small and realistic: five prompts you actually send, with a saved example of how you wrote the same thing yourself. Run the five prompts, put the outputs next to your originals, and check the measurable rules: whether sentence length sits within range, whether the openers match yours, whether every banned phrase is absent, and whether the register shifts are honoured.

What you will usually see is Practice 1 in action. On a precise setup, the new model scores better. On a vague setup, it produces smoother text that is further from you. That result is the clearest argument for doing the setup work once and doing it properly.

What Do the Best ChatGPT Practices Look Like in a Real Workflow?

Set up once, test weekly, maintain monthly. The initial setup takes about an hour if you extract your rules by hand and about ten minutes if you start from a generated profile. After that the maintenance is light.

A working routine looks like this:

Baseline (once). Measurable rules in the "How would you like ChatGPT to respond?" field, structured as ALWAYS / NEVER / CONTEXT SHIFTS.
Context (per project). Documents, glossary and audience notes in a Project. Voice overrides only where the register is genuinely different.
Evidence (once, refreshed yearly). Three to five annotated samples of your real writing, attached where they fit: in custom instructions if short, in the Project if long.
Correction log (weekly). Edits captured as rules, promoted to the baseline when they repeat.
Regression test (per model release). Five fixed prompts compared against your own originals.

None of these steps is a prompting trick. All of them are about what ChatGPT knows before the prompt arrives. That is where the best ChatGPT practices actually live.

How Do You Know Your Setup Is Working?

The output needs fewer edits, and the edits you still make are about substance rather than voice. That is the only measure that matters, and you can track it with the correction log from Practice 6.

Three signs you are there:

You send ChatGPT drafts to colleagues without rewriting the opening and closing lines.
Someone who knows your writing cannot pick the AI-assisted email out of a set of five.
A model update makes the output better rather than different.

Three signs you are not:

Your custom instructions contain the words "professional", "friendly", "engaging" or "natural" without a rule attached to each one.
You are still deleting the same stock phrase every day.
Your voice setup lives in one Project and nowhere else.

If the second list describes your setup, the fix is a profile built from your writing rather than from your description of it. Describing yourself produces the average. Measuring yourself produces you.

Frequently Asked Questions

What are the best ChatGPT practices for professional writing?

Separate what the model can do from what it knows about you, and invest in the second. Write measurable rules (sentence length, openers, banned phrases, register shifts) into custom instructions instead of tone adjectives. Keep subject context in Projects. Log every edit as a rule candidate, and rerun a fixed test set after each model release so you can see whether your setup is being followed more faithfully or drifting.

Do better ChatGPT models make custom instructions unnecessary?

No. Model capability and your own data are separate layers. A new model improves at following instructions, which means it executes a precise profile better than before. It cannot learn how you write a difficult email from training data, because that evidence is in your writing, not in the model. Better models raise the return on a good setup and make a vague one look more polished while sounding less like you.

Why does ChatGPT ignore my tone instructions?

It is not ignoring them. It is following them to the average. "Confident and warm" maps to what confident, warm text looks like across everyone who has written those words, which is nobody in particular. Replace each adjective with a rule the model can check: a sentence length range, a list of openers you use, a list of phrases you never use. Countable rules get followed. Moods get approximated.

How often should I update my ChatGPT setup?

Review your correction log weekly for the first month, then monthly. Rerun your fixed test set after every model release, since interpretation of the same instructions shifts between versions. Refresh your writing samples about once a year, or when your role changes enough that your register has moved. The baseline rules themselves should be stable. If you are rewriting them constantly, they were probably adjectives rather than rules.

What is the difference between ChatGPT custom instructions and a Style Profile?

Custom instructions are a field in ChatGPT where you type rules you can articulate. A Style Profile is the content you put there, extracted from your own writing by measuring it: sentence length distribution, qualifier placement, transition habits, punctuation. You can build one by hand or generate one. Either way, the profile is the measurable rule set, and custom instructions are one of the places it gets deployed, alongside Claude and Gemini.

If you would rather measure your patterns than guess at them, a Style Profile turns your own writing into the rule set above: a Writing Twin for ChatGPT, Claude and Gemini, ready to paste into the field where it belongs. The best ChatGPT practices start with you, not with the prompt.