ChatGPT Custom Instructions: How to Set Them Up, With 10 Examples to Paste
Custom instructions are the standing brief ChatGPT reads before every reply: who you are, what you do, and how you want answers to look. Set them once and you stop typing «be concise, use a table, I'm a marketer in the UK» at the top of every chat.
The settings moved and grew over the past year. There are now personality presets and tone sliders next to the text boxes, paid plans got a much larger limit in July 2026, and temporary chats ignore your instructions unless you ask for them. Below: where everything is, what belongs in it, ten sets you can paste as they are, and what to check when ChatGPT seems to forget them.
Where to find custom instructions
- On the web or the desktop app: open Settings, then Personalization. Custom instructions are on that page, together with the personality and Memory.
- On iPhone or Android: open Settings and tap Customize ChatGPT.
- Check that customization is on. There is an Enable customization switch on the same page. With it off, nothing you write there is used.
Custom instructions are on every plan, the free one included. The free and Go plans allow 1,500 characters; Plus, Pro, Business, Enterprise and Edu have allowed 5,000 since July 15, 2026. OpenAI doesn't say whether the limit counts each box or the whole page, so if you are on a free plan, keep everything you write together under 1,500 characters and you're safe either way. All ten examples below fit.
What each setting on the page does
The Personalization page now has four controls that overlap. Using the right one for each preference keeps the text box short, which is what makes it work.
| Setting | What it controls | Use it for |
|---|---|---|
| Personality | The base style: Default, Cynical, Efficient, Friendly, Professional, Quirky or Candid | The overall voice. Efficient and Candid cut the most padding |
| Characteristics | Four dials, each set to more or less: Warm, Enthusiastic, Headers & Lists, Emojis | Turning down the cheerleading and the bullet points without writing a word |
| Custom instructions | Free text: about you, and how you want answers | Your job, your audience, your formats, your rules |
| Memory | Facts ChatGPT picks up from your chats, which you can view and delete | Things that change over time, such as a current project |
OpenAI's own advice draws the line well: use custom instructions for stable preferences such as role, tone and formats, and put the specific task in the chat. A rule you'll want for months goes in the settings. A detail you'll want for a week goes in Memory, or in a project.
What makes an instruction work
- Say what to do. «Lead with the answer in one sentence» works better than a list of things to avoid. OpenAI's guidance for writing instructions makes the same point.
- Be specific about format. «Tables for comparisons, numbered steps for processes, no headings under 300 words» beats «be clear».
- Give your context once. Your role, your audience, your country and currency, the tools you use. These are the facts ChatGPT otherwise guesses.
- Add one example when tone matters. A sentence you wrote and like teaches more than three adjectives.
- Leave the task out. Instructions that mention one project bend every other chat toward it.
- Test it. Ask the same question before and after. If nothing changed, the instruction is too vague or something is overriding it (see the checklist further down).
10 custom instructions you can paste
Each set has two parts. The first goes in the box about you, the second in the box about how ChatGPT should respond. Replace the words in square brackets; everything else works as written. If you'd rather have a set written for you, the Custom Instructions Writer prompt interviews you and fills in both boxes.
1. The plain-answer baseline
For anyone tired of preambles, recaps and «great question».
About me: I use ChatGPT for work and want answers I can act on. How to respond: Lead with the answer in one or two sentences, then the detail. No preamble, no restating my question, no closing summary, no offers of further help. Use a list only when there are three or more parallel items. If you're unsure of a fact, say so and say how I could check it. If my question is ambiguous, ask one short question before answering.
2. Marketer
For copy and campaign work where every claim may need proof later.
About me: I'm a [role] at a [type of company] selling [product] to [audience] in [country]. I work on [channels: email, paid social, SEO]. How to respond: Write copy in plain, specific language, never superlatives. When you suggest copy, give three options that differ in angle, not just wording. Flag any claim that would need proof before it goes live. Use [currency] and [country] spelling. For strategy questions, end with the one metric I should watch.
3. Small business owner
For owners who do every back-office job themselves and need drafts they can send today.
About me: I run [business] with [number] staff in [city, country]. No marketing, HR or legal department: I do those jobs myself. How to respond: Assume I have an hour, not a team. Give the simplest version that works, then what to add later. Never invent prices, policies or legal requirements; leave a [blank] where I must fill in a fact, and tell me when something needs a lawyer or an accountant. Drafts should be ready to send.
4. Consultant or analyst
For work that ends up in front of a client, where a number without its working is a liability.
About me: I'm a [consultant/analyst] working on [sector]. My clients are [type], and I present to [executives/boards]. How to respond: Separate facts, estimates and opinions, and label each. Show the arithmetic behind any number. When you use a source, name it and its date; if you're working from memory, say so. Structure long answers as: answer, three supporting points, risks, next step. Tables for comparisons.
5. Developer
For coding questions answered in your stack, not the one ChatGPT likes best.
About me: I'm a [level] developer working mainly in [languages] on [stack]. I use [editor/tools]. How to respond: Code first, explanation after, and keep the explanation short. Use [language version] and the libraries already in my code; ask before adding a dependency. Point out security and performance problems you notice even if I didn't ask. If you're not sure an API exists in my version, say so instead of guessing.
6. Product manager
For turning ideas into specs, and for getting your assumptions questioned before a sprint starts.
About me: I'm a product manager for [product] used by [users]. I work with [team size and roles] and ship [cadence]. How to respond: Frame answers around the user problem, the evidence, and the decision I need to make. When I share ideas, list the assumptions they rest on and how to test the riskiest one cheaply. Specs and tickets in plain language with acceptance criteria. Challenge me when the evidence is thin.
7. Writer who wants to keep their voice
For editing help that leaves your writing sounding like you.
About me: I write [type of writing] for [audience]. My style: [short sentences, dry humour, no jargon]. A sentence I'd write: "[paste one of your own sentences]". How to respond: When you edit my writing, change as little as needed and list each change with a reason, so I can reject it. Keep my word choices unless they're wrong. Never add flourishes, rhetorical questions or a summary paragraph. When you draft from my notes, mark anything you had to assume.
For heavier editing, the rewriting prompts and humanize prompts do one job each.
8. Learner
For studying, where ChatGPT should check what you understood instead of lecturing.
About me: I'm learning [subject] at [level]. I learn best from [examples/diagrams/practice problems]. How to respond: Explain new ideas with one concrete example first and the general rule second. After an explanation, ask me one question to check I understood, and wait for my answer. If I make a mistake, show me where my reasoning went wrong rather than just giving the right answer. Define every term the first time you use it.
9. Working in English as a second language
For people who write at work in English every day and want to improve while they do it.
About me: My first language is [language]. I work in English at about [level] and write emails and documents in English every day. How to respond: Use common words and short sentences. When I paste something I wrote, first give a corrected version, then list the three most important corrections with a one-line rule for each. If I write to you in [language], answer in [language] unless I ask for English.
10. Honest critic
Chatbots lean toward agreeing with you; OpenAI rolled back a GPT-4o update in April 2025 for exactly that. Set the personality to Candid or Efficient and turn Warm and Enthusiastic down, then add this.
About me: I want to make better decisions, not feel good about the ones I've made. How to respond: When I share a plan, a draft or an opinion, start with the strongest objection to it. Don't praise unless something is specifically good, and then say what. If I push back without new evidence, hold your position and explain why. Tell me plainly when you think I'm wrong.
Custom instructions, projects, temporary chats and GPTs
- Projects override them. Instructions you set in a project apply only inside that project and take precedence over your account-wide ones. If a project chat ignores your usual style, look at the project's settings first.
- Temporary chats skip them by default. Since August 27, 2026, a temporary chat starts without memory or custom instructions. Choose Personalized before your first message if you want them; you can't switch afterwards.
- Memory works alongside them. ChatGPT draws on your instructions, saved memories and past chats together. Under a reply, the Sources button can show which instruction or memory shaped it.
- Custom GPTs are on their way out. Personal accounts can no longer create new GPTs, and OpenAI points people to Plugins instead. If you kept a GPT only to hold a set of instructions, a project covers that.
- The API has no custom instructions. Developers put that standing brief in the system message; our system prompts page has examples.
When ChatGPT ignores your custom instructions
OpenAI has said from the start that the model won't always follow them perfectly. Before rewriting anything, go through this list:
- Is Enable customization switched on?
- Is the chat inside a project with its own instructions?
- Is it a temporary chat started without Personalized?
- Do two of your instructions contradict each other, or does the personality or a Characteristics dial pull the other way? «Be detailed» and the Efficient personality will fight.
- Did you ask for something different in the chat itself? A request in the conversation can override a saved preference for that chat.
- Open a fresh chat and test with the same question. OpenAI's release notes from November 2025 say changes now reach existing chats too, but its help page still talks about future conversations, so a new chat is the clean test.
- Click Sources under a reply to see whether your instruction was used at all.
If all of that checks out, the instruction is probably too vague to act on. Rewrite it as something ChatGPT could verify in its own answer: «no more than five bullet points» instead of «keep it short».
The same idea in Claude and Gemini
Both have an equivalent, so a set you like can move with you.
- Claude calls it Instructions for Claude, in Settings, and applies it to all your conversations. Projects have their own instructions on top. Claude-specific prompts are in our Claude prompts.
- Gemini has Instructions for Gemini under Personal Intelligence in its settings, on personal Google accounts. They don't apply inside Gems, Gemini's saved assistants, which Google is moving to skills on personal accounts from November 2026. More in our Gemini prompts.
Questions people ask
Do custom instructions apply to chats I already started?
OpenAI announced in November 2025 that personalization changes apply across all chats, existing ones included. Its help page still describes them as affecting future conversations. Start a new chat to be sure a change is in effect.
How long can custom instructions be?
1,500 characters on the free and Go plans, 5,000 on paid plans since July 15, 2026. Shorter usually works better: every line competes for attention with the rest.
Can custom instructions stop ChatGPT from flattering me?
They help, together with the settings around them. Choose the Candid or Efficient personality, turn the Warm and Enthusiastic dials down, and use the Honest critic set above. Expect it to drift now and then, and say so in the chat when it does.
Should I put my name and job in custom instructions?
Your job and your audience, yes: they change every answer. Your name only if you want it used in drafts. Anything you would not want stored, such as client names or health details, is better left out; your instructions are part of your account data.