ChatGPT Custom Instructions That Actually Help

Custom instructions are the highest-value setting in ChatGPT and the one most people fill in once, badly, and never revisit. Done well, they stop you re-explaining who you are in every conversation. Done the usual way — a paragraph of adjectives — they change almost nothing.

This guide covers what actually belongs in each field, what to keep out of them, and why the thing you want may belong in memory instead.

Where to Find Them

Open your profile menu and choose Personalization (labelled Customize ChatGPT in some versions), then make sure customisation is switched on. Settings apply to new conversations — an existing chat will carry on behaving the way it already was, which is the first thing that confuses people testing changes.

The Four Fields, and What Belongs in Each

The setting used to be two boxes and is now four. Each answers a different question, and putting the right thing in the right box matters more than writing a lot.

Write Constraints, Not Adjectives

This is the difference between instructions that work and instructions that decorate. «Be professional and helpful» describes something the model was already trying to do, so it changes nothing. A constraint removes a behaviour you keep having to correct.

Compare what people usually write with what actually shifts the output:

A useful test: read your instruction and ask whether ChatGPT could plausibly have done the opposite. If not, the line is doing no work.

Custom Instructions or Memory?

These two features overlap and are constantly confused. The split is simple once you see it.

Put things that will still be true in a year into instructions. Let memory carry what is true this month. If you find yourself editing instructions weekly, that content probably belongs in memory or in a project instead.

A Starting Set Worth Copying

Adapt rather than paste — the value comes from the constraints matching your actual irritations.

• Traits: «Be direct. No preamble, no summarising my question back to me. If a question is ambiguous, ask before answering.»
• Traits: «Tell me when you're uncertain or when something is outside your knowledge, rather than producing a confident guess.»
• Anything else: «Default to short answers. Expand only when I ask for depth.»
• Anything else: «When you give code, no line-by-line explanation unless I ask — just the code and anything genuinely surprising in it.»

What to Keep Out

The box is not a prompt library, and length works against you. Every instruction competes for attention with every other one, so a long list dilutes the ones that matter.

• Project-specific detail — that belongs in a project or a chat, not in every conversation you ever have.
• Long lists of forbidden words. Negative instructions are followed inconsistently, and each one you add weakens the rest.
• Anything sensitive. This text is attached to your conversations; treat it as you would a shared document.
• Contradictions, like asking for brevity and thorough explanations in the same breath.

Common Mistakes to Avoid

• Testing changes in an existing chat — instructions only apply to new conversations.
• Writing adjectives instead of constraints, then concluding the feature does nothing.
• Filling all four boxes because they are there, rather than because you have something to say.
• Setting it once at signup and never revisiting, when your work has since changed.
• Using instructions for context that changes weekly, which memory handles better.

Where to Go Next

If the vocabulary around prompts and context is still fuzzy, the ChatGPT glossary explains the terms plainly, and the ChatGPT knowledge base covers how the assistant works underneath. For plans, features and honest limitations, see the ChatGPT tool page.

Bottom Line

Good custom instructions are short, specific and made of constraints you can check. Write down the corrections you keep repeating, turn each into one line, and leave everything else out. Then revisit in a month — the instructions worth keeping are the ones you never think about, and the ones you keep fighting are wrong.