How to Set Up ChatGPT Custom Instructions: Get Better Responses Every Time

Most complaints about ChatGPT are complaints about a setting nobody opened.

Responses too long. Too much preamble before the answer. Explaining basics you already know. Hedging every claim into uselessness. Opening with "Great question!" every single time. These read like model limitations and they are almost entirely configuration — fixable permanently, in about ten minutes, in a settings panel most people have never visited.

I spent months re-typing "be concise, I'm technical, skip the preamble" at the start of conversations before finding it. That is the whole pitch: say it once instead of every time.

What They Actually Are

Two text boxes, roughly 1,500 characters each, applied automatically to every new conversation.

The first holds context about you — your work, your field, what you are usually trying to do. The second holds instructions about output — length, tone, structure, and what you never want to see.

The split matters more than it looks. The first box stops the model explaining your own profession back to you. The second stops it writing in the default register, which is the one everyone finds slightly irritating and nobody bothers to change.

To find them: profile picture, bottom-left, then Settings → Personalization. On mobile it is the menu icon, then Settings, then Personalization. OpenAI has relocated this menu more than once, so if the path has moved again, search the settings for "personalization" rather than "custom instructions".

Box One: Context

Be specific about the things that change how an answer should be written. Your field, your seniority within it, the tools you actually use, and who your work is for.

Marketing:

I'm a marketing manager at a B2B SaaS company. Our audience is small business owners who aren't technical. I work mainly on email campaigns, landing pages, and social content.

I'm comfortable with marketing terminology but prefer straightforward explanations over jargon. We use HubSpot for CRM and Mailchimp for email.

Development:

I'm a full-stack developer working mainly with React, Node.js, and PostgreSQL, using TypeScript in most projects. My code runs in production, so I care about error handling and edge cases.

I don't need basic syntax explained. I deploy to AWS using Docker.

Study:

I'm a second-year computer science student. I understand programming basics but I'm still learning advanced concepts.

I need to understand things deeply for exams, not just get working answers. I learn best from examples and analogies.

Writing:

I'm a freelance content writer covering technology and business. I write for American audiences and follow AP style.

I value clarity over cleverness. My pieces run 1,500-2,500 words and need to be readable, not keyword-stuffed.

Notice what these have in common: every line changes what a good answer looks like. "I like technology" would not. If a sentence would not alter a response, it is taking up characters you need elsewhere.

Box Two: Output

This is where the quality change happens, and the rule is narrow enough to state in one line: write constraints, not aspirations.

"Be helpful" does nothing — it was already trying. "Do not open by restating my question" is a specific behaviour it can act on. Nearly every set of instructions that fails, fails on this distinction.

Concise:

Be direct. Start with the answer, then explain if needed.

Skip preamble like "Great question!" or "I'd be happy to help." Keep paragraphs to 2-3 sentences.

If I ask for a recommendation, give me a clear stance rather than "it depends" — unless it genuinely does, in which case tell me what it depends on.

Don't add disclaimers unless there's a real accuracy or safety concern.

Learning:

Explain step by step. Use analogies connecting new ideas to things I already know.

After explaining, give me a practical example. If my reasoning is wrong, correct me directly and say why.

Code:

Provide working code, not pseudocode, with error handling included.

Comment the non-obvious parts. If there are multiple approaches, mention the alternatives briefly. Use current syntax — don't suggest deprecated methods.

Editing:

When I share a draft, focus on what could be improved rather than praising what works.

Give specific suggestions, not "make it more engaging." If you suggest a change, show me how it would read.

Don't rewrite the whole piece — address what I asked about.

The single instruction I would recommend to everyone regardless of field: tell me when you are uncertain rather than guessing. It changes output more than any other line, because the default behaviour is confident fluency whether or not confidence is warranted, and that is the failure mode that actually costs you something.

The Four Ways This Fails

Aspirations instead of constraints. Covered above, and it is the big one. "I want high-quality responses" is not actionable. "Use tables when comparing three or more things" is.

Overloading the boxes. You have 1,500 characters and the temptation is to fill them. Instructions that apply to five percent of your conversations dilute the ones that apply to all of them. Write for the typical case.

Contradicting yourself. "Be concise" alongside "explain everything thoroughly" produces something arbitrary. If you need both, make it conditional: be concise by default; when I say "explain" or "why", go into detail.

Setting and forgetting. Instructions written while you were learning a subject will be wrong once you know it, and you will experience that as the model becoming annoying rather than as your own configuration going stale. Revisit every few months.

Conditional Instructions

Once the basics are working, conditional rules are the upgrade worth knowing:

For coding questions: working examples with comments.
For writing help: structure and flow over grammar.
For research: cite sources and flag uncertainty.
For comparisons: use a table when there are 3+ items.

You can also name your stack, which removes a recurring round of clarification:

For project management, assume Notion.
For repos, GitHub with conventional commits.
When suggesting tools, prefer options that work offline.

Testing It

Do not assume it worked. Run four checks: ask something in your field and see whether the terminology level is right; ask for an explanation and check the depth; ask for a recommendation and see whether the format matches; and read the tone.

If something is off, the fix is almost always adding a constraint rather than rewriting everything. Adjust one line, test again.

Against Custom GPTs and Memory

These three overlap and get conflated constantly.

Custom instructions are global and deliberate. Every conversation, always, exactly what you wrote.

Custom GPTs are separate assistants with their own instructions and uploaded knowledge, built for one repeated job — a blog assistant, a code reviewer. Worth it when a workflow recurs enough to justify the setup. The custom GPTs guide covers building one.

Memory is what ChatGPT picks up by itself and keeps. Useful when it retains something relevant, irritating when it retains something situational and quietly applies it for weeks. Unlike instructions, you did not choose it — so review it occasionally. ChatGPT tips and tricks covers pruning it.

Most people who get real value use instructions for standing preferences and a GPT or two for genuinely repeated workflows.

Why This Beats Prompting Advice

There is a large industry teaching prompt formulas. Most of it is compensating for context the model does not have — and standing context belongs in a setting, not retyped every session.

Set this up and a good deal of prompting advice stops being necessary, because the model already knows who it is talking to. The techniques that still matter after that are covered in writing better prompts, and there are fewer of them than you would expect. For applying any of this professionally, how to use ChatGPT for work covers where the output is trustworthy and where it is not.

Ten minutes, once. It is the highest-return configuration change available in the product, and the reason it goes unused is simply that nothing prompts you to open the panel.