Writing ChatGPT Custom Instructions That Actually Stick
Open Custom Instructions, stare at two empty boxes, and most people type a sentence or two, save it, and forget about it. Then they wonder why ChatGPT still asks "what's the context here?" three messages into a new chat. The instructions were too thin to do anything. This is the setting that's supposed to eliminate that repetition, but only if you actually fill it in like it matters.
If you haven't set up Custom Instructions at all yet, the Complete Beginner's Guide to ChatGPT walks through where to find the setting. This article is about what to actually write once you're there.
What the two boxes are really asking
Custom Instructions centers on two questions, and the framing matters more than people give it credit for. (ChatGPT's Customize dialog may also show extra fields such as a name, an occupation, and personality traits, but the two free-text questions below are where most of the value is.)
The first box asks what ChatGPT should know about you to give better responses. This isn't a biography. It's the facts that change how an answer should be shaped: your job, your industry, the tools you use daily, your experience level in the topics you ask about most, and anything you'd otherwise have to explain in every single conversation.
The second box asks how you want ChatGPT to respond. This is about behavior, not background: tone, format, how much hand-holding you want, whether you want opinions or just options, how long answers should run by default.
The mistake people make is treating both boxes as one undifferentiated "about me" field. A sentence like "I'm a busy marketing manager who likes clear answers" technically touches both boxes but does neither job well. It doesn't tell ChatGPT anything specific enough to change its output.
Note
Custom Instructions apply to your new chats by default, and you can switch them off in settings. They're a standing default, not a one-time prompt, which is exactly why vague ones are a wasted setting.
The three mistakes that make instructions useless
Too vague. "I like concise, helpful answers" describes almost everyone who has ever used ChatGPT. It gives the model nothing to grab onto. Compare that to "Cap responses at 150 words unless I explicitly ask for more detail, and skip the recap of what I just asked you." The second version is a rule ChatGPT can actually follow.
Too long. The opposite failure is a paragraph that reads like a personal manifesto, covering your career history, your five favorite frameworks, and a list of pet peeves. Long instructions dilute themselves. ChatGPT weighs every sentence you give it, and a wall of text buries the two or three details that actually change behavior in a pile of ones that don't. Four to six tight sentences per box outperforms a page every time.
Contradicting itself. This one is sneaky. "Be brief" in one line and "always explain your full reasoning before giving an answer" in another line will fight each other, and you'll get inconsistent results depending on which instruction the model weights more heavily for a given question. Read your own instructions back and ask whether two of your sentences could both be followed at once. If not, cut one.
Three real, filled-out examples
A freelance copywriter
What should ChatGPT know about you: "I'm a freelance copywriter working mostly with B2B SaaS clients. I write website copy, email sequences, and LinkedIn posts. I know marketing terminology well, so don't explain basic concepts like 'value proposition' to me. My clients skew mid-size (50-500 employees), not enterprise or scrappy startups."
How should ChatGPT respond: "Give me copy in my requested format immediately, don't ask clarifying questions unless something is actually ambiguous. Default to a direct, confident tone, not overly friendly or exclamation-heavy. When you give me multiple options, label them by strategy (e.g. 'urgency-led' vs 'social-proof-led') instead of just numbering them."
A high school biology teacher
What should ChatGPT know about you: "I teach 9th and 10th grade biology at a public high school. My students are 14-16 years old with a wide range of reading levels. State standards require me to tie assessments to specific learning objectives, not just general topic coverage."
How should ChatGPT respond: "When I ask for quiz questions or explanations, keep vocabulary at a 9th-grade reading level unless I say otherwise. Always ask which learning objective a request should map to if I haven't specified one. Format quiz questions with the answer key separated at the bottom, not inline."
A solo operations lead at a small logistics company
What should ChatGPT know about you: "I run operations for a 12-person logistics company. I handle vendor contracts, scheduling, and internal process docs. I'm not a lawyer, so flag anything in a contract that sounds like it needs one, don't just summarize it as fine."
How should ChatGPT respond: "Default to bullet points over paragraphs for anything process-related. When summarizing a document, always separate 'what this says' from 'what I'd want to double-check.' Don't soften bad news about a vendor or a delay, just tell me plainly."
Notice what these three have in common: specific job context that changes the actual content of an answer, and specific behavioral rules that change the actual shape of an answer. Neither box is doing the other's job.
What should ChatGPT know about you: I manage a 6-person customer support team at a mid-size e-commerce company. I write internal SOPs, handle escalations, and review support scripts. I'm comfortable with technical language around our ticketing system (Zendesk) but not with marketing jargon.
How should ChatGPT respond: Write in plain, direct language, no corporate filler like "circle back" or "synergize." When drafting an SOP, always include a numbered steps section and a separate "common exceptions" section. If I paste an escalation, tell me the likely root cause before suggesting a response.
”Two more professions worth seeing filled out, since the shape of a good instruction set changes with the actual job, not just the industry.
A solo real estate agent
What should ChatGPT know about you: "I'm an independent real estate agent working residential sales in a mid-size suburban market. Most of my clients are first-time buyers who don't know real estate terminology yet. I write listing descriptions, buyer follow-up emails, and social media posts."
How should ChatGPT respond: "When writing anything client-facing, keep language plain and explain a term like 'contingency' or 'escrow' in the same sentence if you have to use it at all. For listing descriptions, lead with the detail most likely to matter to a first-time buyer, like the school district or the commute, not generic praise like 'stunning' or 'must-see.' Keep social posts under 80 words."
An engineering manager at a mid-size startup
What should ChatGPT know about you: "I manage a team of seven engineers at a 60-person startup. I write performance reviews, sprint retro notes, and feedback on technical design docs. I have a strong technical background myself, so don't simplify architecture discussions for me."
How should ChatGPT respond: "For performance review drafts, always separate 'what they did' from 'the impact on the team,' don't blend the two into one sentence. When reviewing a design doc, lead with the biggest structural risk before commenting on smaller issues like naming. If you spot a flaw in my own reasoning, say so directly instead of agreeing by default."
A quick reference: what goes in each box
If a specific detail still feels like it could go either way, this is the split that actually holds up across all five examples above:
| Type of detail | Box 1: what ChatGPT should know | Box 2: how it should respond |
|---|---|---|
| Job, industry, tools you use daily | Yes | No |
| Your existing expertise on the topic | Yes | No |
| Who your output is usually for (clients, students, your own team) | Yes | No |
| Tone: direct, formal, casual | No | Yes |
| Length or format defaults | No | Yes |
| Whether to ask clarifying questions or just proceed | No | Yes |
| How to handle disagreement or a spotted mistake | No | Yes |
The pattern: Box 1 answers "what's true about my situation." Box 2 answers "what should this response look like, regardless of the topic." A detail that would still matter no matter what you were asking about belongs in Box 2. A detail that's only relevant because of who you are or what you do belongs in Box 1.
How this interacts with Memory (and why they're not the same thing)
This is where a lot of people get confused. Custom Instructions and Memory both feel like "ChatGPT remembers me," but they work differently and serve different jobs.
Custom Instructions is a fixed brief you write once and edit deliberately. It doesn't change on its own. It applies the same way to every new chat until you go back and edit it yourself.
Memory is dynamic. ChatGPT saves specific facts as you chat (sometimes because you told it to remember something, sometimes because it inferred something worth keeping) and those saved memories accumulate and evolve over time without you writing them into a settings box. For the mechanics of what memory saves automatically versus what you have to say explicitly, see what ChatGPT's memory actually remembers.
The practical difference: put durable, stable facts in Custom Instructions (your job, your standard format preferences, your industry). Let Memory pick up the things that shift chat to chat (a project you're currently working on, a preference you mentioned once in passing). If you find yourself repeatedly correcting a memory that keeps resurfacing wrong, that's a sign that fact belongs as an explicit line in Custom Instructions instead, where it's stable and you control it directly.
A quick test
If a fact about you would still be true in a year, it belongs in Custom Instructions. If it's only true this week or this project, let Memory handle it, or mention it in the chat itself.
Revisit it after a month
The instructions you write on day one are a guess. After a few weeks of real use, you'll notice patterns: ChatGPT keeps over-explaining something you already know, or keeps under-formatting something you always want structured a certain way. Go back and add one sentence addressing that specific friction rather than rewriting the whole thing. Custom Instructions works best as a document you tune over time, not a form you fill out once and never open again.
Official sources
Checked on September 21, 2026. Features, plans and names change often, so the vendor's own pages are the final word.