ChatGPT custom instructions: what to actually write

Custom instructions are a block of text applied to every new conversation. In ChatGPT they sit under Settings → Personalization → Custom Instructions, usually split into two questions: what should ChatGPT know about you, and how should it respond. Claude and Gemini have equivalents under personal preferences and saved info.

Field one: what it should know about you

Keep it to practical, general facts that change the substance of answers. A good version reads like a note you'd leave a competent house-sitter:

• Two adults and one school-age child; I do most of the cooking and scheduling.
• I work weekdays and have roughly 30 minutes for weeknight dinners.
• We don't eat pork; one of us dislikes very spicy food.
• No car — plans should assume public transport or walking.
• Right now I'm trying to get on top of household admin and read more.

Note what's absent: no names, no address, no employer, no financial or medical details. None of that would improve a single answer.

Field two: how it should respond

This is where most of the quality improvement comes from. A strong default set:

• Be concise and lead with the answer. No preamble or restating my question.
• Use short paragraphs or bullets. Avoid headings for short answers.
• When something is genuinely uncertain, say so plainly and tell me what to verify instead of guessing specifics like prices, hours, or availability.
• If my request is ambiguous, ask one clarifying question before answering at length.
• Give me two or three options with tradeoffs when there's a real choice to make.
• Skip motivational language, disclaimers I didn't ask for, and flattery.

Common mistakes

Writing an essay

Long instructions dilute themselves. Under 200 words per field is the sweet spot; the model follows short, concrete rules far more reliably than long prose.

Vague adjectives

"Be helpful and professional" changes nothing. "Under 150 words unless I ask for more" changes everything. Prefer testable rules.

Contradictions

Asking for brevity and thorough detail in the same breath produces inconsistent results. Pick a default and override it in the conversation when you need the opposite.

Letting it go stale

Instructions written a year ago quietly distort answers. Reread them when your circumstances change.

Test whether it's working

Ask three questions you'd normally ask — a meal idea, a short email, a comparison — and check the format and assumptions. If they don't reflect your instructions, make the relevant rule more specific and blunt. It's normal to need two or three revisions.

Instructions versus memory

Instructions are stable rules you control. Memory is what the assistant picks up as you chat — useful, but less predictable. Keep the things you care about in instructions, and review memory periodically, as described in making ChatGPT remember your preferences.

Not sure what your instructions should optimize for? Start with the free AI assessment.

For a ready-made version of this text, the Personal AI Profile builder generates a short profile sized for the custom-instructions box.