Best ChatGPT Custom Instructions: 4 Instruction Types Compared

TL;DR: The best ChatGPT custom instructions are the ones ChatGPT can execute, and "professional yet friendly" is not executable. This post compares the four instruction types people actually paste into the responding field: tone labels, rule lists, pasted samples and a measured Style Profile. Labels describe thousands of people. Rules fix a handful of habits. Samples force ChatGPT to guess which patterns are yours. A profile built from your real writing hands it the numbers. Put the profile in the field, add a short rule list under it, and skip the adjectives.

Search for recommended ChatGPT custom instructions and you get lists of templates. Paste one in and the output improves for about a week, then you notice it reads like everyone else who pasted the same template.

The template was never the problem. The type of instruction was.

What Are the 4 Types of ChatGPT Custom Instructions?

Every custom instruction set for writing style falls into one of four types, defined by what the text actually contains: adjectives, rules, examples or measurements. The container (global settings, a Project, a custom GPT) changes where the instructions apply. The type changes whether ChatGPT can follow them at all.

The field itself is the same in every case. OpenAI's custom instructions guide describes two boxes: a short one about you and a longer one for how ChatGPT should respond. Our character limit breakdown covers what fits where. Here we care about what goes in.

Most templates on the web are type 1 with a little type 2 mixed in. That is why they all sound the same.

Why Does a Tone Label Fail as a Custom Instruction?

Because AI does not know your tone; it knows words that describe tone. When you type "professional yet friendly", ChatGPT maps those words to the register it has learned for them, and that register is identical for every user who types the same words. The label is an average, and you are asking the model to be average on purpose.

Try the test yourself. Write "professional yet friendly" into a fresh conversation and ask for a follow-up email after a meeting. Then ask a colleague to do the same. The two emails will be closer to each other than either is to anything you actually sent last month. The label carried no information that separates you from your colleague, so the output could not either.

A voice lives somewhere the label cannot reach:

• Sentence length. Not "concise" but a distribution: how short your short sentences get, how long the long ones run, and when you switch.
• Qualifier placement. Whether "probably" sits at the start of the clause or the end, and whether you hedge at all when you have decided.
• Transitions. "So" versus "therefore" versus nothing. Whether a paragraph opens with the conclusion or builds to it.
• Punctuation. Colon rate, comma rate, whether a parenthesis ever appears, whether you use one dash or none.

None of those is an adjective. A label cannot encode any of them, and a model cannot infer them from one. If your current instructions are mostly adjectives, that is the whole explanation for why the output drifts to generic.

Why Do Rule Lists and Pasted Samples Fall Short?

Rule lists fix the habits you can name, and pasted samples cover the ones you cannot, but neither gives ChatGPT a target it can measure itself against. Both are real improvements over a label. Both also hit a ceiling that a lot of people mistake for the limit of what ChatGPT can do.

Rule lists: strong on constraints, blind on rhythm

A rule list is the right tool for anything binary. Banned phrases, sign-off format, whether to use headings, whether to bold. ChatGPT follows these well because each rule has a yes or no answer. Our six ready-to-use templates are built this way, and they are a fine scaffold.

The ceiling is that rhythm is not binary. "Vary sentence length" is a rule ChatGPT will nod at and then produce the same medium-length sentence twelve times, because it has no number to aim for. "Keep it tight" gets you the same thing. You can write a hundred rules and still never describe the shape of a paragraph you would write, because the shape is a pattern, and rules describe events.

Pasted samples: better evidence, unreliable inference

Pasting three emails and saying "write like this" is the first type that gives ChatGPT actual evidence of how you write. It works noticeably better than adjectives. Where it breaks is the inference step. ChatGPT reads three emails and has to decide which features are your voice and which are accidents of those three emails. Was the short opener a habit or a Tuesday? It cannot know, and it will guess differently in every conversation.

The other limit is volume. Three emails cannot tell anyone your average sentence length across a year of writing; the sample is too small for the patterns that matter. Project knowledge files raise the ceiling somewhat, and we measured where that ceiling sits. It sits below a profile, for the same reason: the inference happens fresh every time, and it is never checked.

What Makes a Measured Profile the Best Custom Instruction?

A measured profile turns your writing into targets ChatGPT can hit and a check can verify. Instead of "concise", it says "mean sentence length 13 words, a quarter of sentences under 8". Instead of "friendly", it says "first line is the decision, greeting second or not at all". Instead of "vary the rhythm", it gives a burstiness range. Every line is something a model can execute and something a script can measure.

That is the difference between prompted and trained. A prompt describes how you hope to sound. A profile built from your own writing is evidence of how you actually write. Here is what the same voice looks like across the four types, for one dimension:

That last row is the one people get backwards. A better model does not make a personal profile less necessary. Model capability and user-data conditioning are separate layers. OpenAI improves the first with every release and cannot supply the second, because your emails are not in the model and never will be. The better the model gets at following nuanced instructions, the better it executes your profile too. A vague label wastes that new capability. A precise profile spends it.

How the profile gets built is covered in how style extraction works. The short version: you add writing you actually sent, the analysis measures the patterns above across all of it, and you get a Master Prompt in the system-prompt format the responding field expects. The first use case (one language, one medium, one audience) is free.

One honest limit. A profile constrains rhythm, vocabulary, punctuation and structure. It carries no facts and no opinions. ChatGPT will write the paragraph the way you write paragraphs, and you still decide what the paragraph says.

How Should You Set Up ChatGPT Custom Instructions Today?

Put a measured profile in the responding field, a short rule list under it, and nothing in the way of adjectives. Here is the layout that works, and it fits the 8,000-character field with room to spare.

1. About you (1,500 characters). Role, audiences, languages. Facts only. This field is for context, not voice.
2. How would you like ChatGPT to respond (8,000 characters). Paste the profile export first. Under it, a rule list of no more than ten hard constraints: banned openers, sign-off, formatting defaults. Delete every rule that describes ChatGPT's own defaults ("be clear", "be helpful"); they cost characters and change nothing.
3. Projects for register shifts. If your client email and your internal notes are different voices, the global field holds the default and a Project holds the other. A profile is a matrix of language, audience and medium, and you can export a single cell of it (for example your client-email mode) straight into the Project instructions. Project instructions override the global ones inside that Project.
4. Test with a check, not a reread. Ask for a real draft, then compare it against the profile's numbers. If ChatGPT is connected to the MyWritingTwin app, it can run the check itself and return a pass or fail per constraint. We built that check as real code because an assistant's own opinion of whether a draft matches is not trustworthy; we watched assistants report passing numbers that the measurement contradicted.

If you write in more than one language, do the profile per language. Register shifts across keigo in Japanese, tu and vous in French, Sie and du in German. AI treats all languages the same and misses those shifts, and a profile built from English email will not carry your Japanese voice.

What to skip: the "About you" field as a voice dump, adjective stacks anywhere, and rules you cannot verify. If you cannot tell whether a draft obeyed an instruction, ChatGPT cannot either.

What to Do Next

1. Open your current custom instructions and sort each line into a type. If most lines are adjectives, that is the fix.
2. Build a free Style Profile from writing you actually sent, and paste the export into the responding field.
3. Keep at most ten hard rules under it. Cut the rest.
4. Run one real draft and check it against the numbers, either by hand or through the app.

The instruction type is the whole game. A Writing Twin for ChatGPT, Claude or Gemini is one profile built once from your real writing, and the responding field is simply where ChatGPT reads it.