ChatGPT Custom Instructions let you set reusable preferences for how the assistant should respond, so you do not have to repeat the same guidance in every conversation. You can describe your background, preferred level of detail, writing style, or how you want uncertainty handled. The feature is most useful when you want a consistent starting point across unrelated chats—not when you need to give detailed instructions for only one task.
Custom Instructions are guidance, not a guarantee. ChatGPT can still misunderstand a request or need more context in the current conversation. The best setup is short, specific, easy to maintain, and paired with a clear prompt for the task at hand. OpenAI’s official Custom Instructions FAQ documents the current controls and limits.
What Custom Instructions do
Custom Instructions are explicit directions about what ChatGPT should know about you and how you prefer it to answer. For example, you might explain that you work in accounting, are learning Python, or need technical explanations for a non-specialist audience. You can also request concise answers, structured headings, examples, or a clear distinction between confirmed facts and assumptions.
Think of them as a default briefing. They help set expectations, but they do not replace the details of a specific assignment. If you want a report comparing three products, you still need to identify the products, decision criteria, and intended audience in that chat.
OpenAI documents Custom Instructions as available on the web, desktop, iOS, and Android. Labels and layouts can change, and workspace administrators may control which personalization features are available.
How to set them up
Web and desktop
- Open ChatGPT and select your profile or account menu.
- Choose Settings, then Personalization.
- Select Custom Instructions.
- Make sure Enable customization is on.
- Enter your preferences in the available fields.
iPhone and Android
- Open the ChatGPT app and go to Settings.
- Choose Customize ChatGPT.
- Turn on Enable customization.
- Enter your instructions and return to the chat.
These are the paths documented by OpenAI at the time of writing. If a label differs in your account, check the current Custom Instructions help page. If the feature appears unavailable in a managed work account, your organization may have restricted a setting or may use a different workspace configuration.
What should you write?
A strong instruction describes a stable preference rather than a one-time task. Useful categories include:
- Background: your role, experience level, language, or recurring work context.
- Audience: whether answers are for a beginner, a technical team, or an executive.
- Style: concise or detailed, formal or conversational, with or without specialist terminology.
- Structure: headings, numbered steps, or tables when they make a comparison easier to scan.
- Evidence: when to link primary sources, distinguish fact from interpretation, and flag uncertainty.
- Limits: what the assistant should not assume or present as verified.
Avoid adding every preference you have ever mentioned. Prioritize rules that improve many different conversations. OpenAI’s prompting best practices also emphasize clarity, specificity, and iterative refinement.
A practical starter template
Adapt this example to your own work:
About me: I work in [field] and usually need answers for [audience or purpose]. Assume I know [baseline knowledge], but explain unfamiliar technical terms.
How to respond: Start with the direct answer, then give the reasoning and practical steps. Use clear headings and concise paragraphs. Use a table only when it makes a comparison easier to understand.
Accuracy: Separate verified facts from estimates or interpretation. For current claims, prefer primary sources and link to them. If evidence is incomplete, state the uncertainty rather than filling gaps.
Interaction: Ask a clarifying question when missing information would materially change the answer. Otherwise, state a reasonable assumption and continue.
Replace the bracketed sections with real context and remove any line that does not suit you. A short instruction you can maintain is often more useful than a long list of rules that repeat or contradict one another.
Five ways to improve your instructions
1. Describe observable behavior
“Be better” and “give high-quality answers” are too vague. “Use plain English, explain specialist terms on first use, and include a worked example for difficult concepts” is actionable. Instead of merely asking for accuracy, ask the assistant to link important claims to reliable sources and identify uncertainty.
2. Set flexible defaults
A rule such as “always be brief” may conflict with a request for a detailed guide. A better version is: “Prefer concise answers, but provide the depth needed to answer the question completely.” It sets a priority without forcing every task into the same shape.
3. Keep one-off requirements in the prompt
A particular article’s word count, a project’s deadline, or a spreadsheet’s column names belong in the relevant chat or project. Mixing temporary details into global instructions makes them harder to maintain and can cause irrelevant preferences to carry into unrelated work. For task-level examples, see our guide to writing better ChatGPT prompts.
4. Explain what to do when information is missing
You could write: “If a critical detail is missing, ask one focused question; otherwise, state your assumption.” This can reduce unnecessary back-and-forth while making important uncertainty visible.
5. Test different kinds of requests
After saving your instructions, try a factual question, a writing task, and a technical request. Check the tone, structure, and level of explanation. If a preference is ignored, make it shorter or more concrete, remove contradictions, and test again. Custom Instructions influence responses; they do not ensure every preference will be followed in every situation.
Know the character limits
OpenAI’s current FAQ lists these maximums:
- Free and Go: up to 1,500 characters.
- Plus, Pro, Enterprise, Business, and Education: up to 5,000 characters.
Limits and product controls can change, so consult the official FAQ if the interface rejects your text or your plan changes. If you approach the limit, remove repeated wording and examples that are not broadly useful. Keep the few rules that most affect answer quality.
Custom Instructions vs. Memory vs. Projects
These features can all influence an answer, but they serve different purposes.
Custom Instructions are the explicit defaults you write yourself: how ChatGPT should respond and what general context it should consider.
Memory can use relevant details from conversations and other available sources to personalize future chats. It is not a complete transcript or a guarantee that every detail will be remembered. Read our guide to how ChatGPT Memory works for the controls and deletion distinctions.
Project instructions apply within a particular workspace for ongoing work. OpenAI says they apply only inside that project and override global Custom Instructions. Projects are a better home for rules that apply to one client, research effort, course, or writing project but would be distracting elsewhere. See how to use ChatGPT Projects.
A practical division is simple: use Custom Instructions for stable response preferences, Memory for relevant remembered context when enabled, Project instructions for work-specific rules, and the current prompt for the immediate task.
Privacy and data considerations
Treat Custom Instructions as account information, not as a secure vault. Keep secrets and confidential client material out of reusable instructions. OpenAI’s FAQ says Custom Instructions are included in your ChatGPT data export.
OpenAI also notes that information from use of Custom Instructions may be used to improve model performance. For personal accounts, review Settings → Data Controls and the Data Controls FAQ to understand whether eligible conversations may be used for model improvement and how to change that setting. Turning off Improve the model for everyone does not delete chat history; it changes the use of new conversations for model improvement. Managed workspaces may have different data practices and administrator controls.
Editing or removing an instruction affects future responses; it does not rewrite messages already in your history. If an older conversation contains information you no longer want there, review the conversation and the relevant account controls separately. For more context, see Are ChatGPT conversations private?.
Troubleshooting
If ChatGPT does not seem to follow your preferences:
- Confirm customization is enabled. Reopen Personalization or Customize ChatGPT and check the toggle.
- Remove conflicting rules. “Always answer in one sentence” conflicts with “explain every step in detail.” Decide which preference takes priority.
- Make the instruction testable. Replace “write professionally” with concrete guidance about tone, structure, and audience.
- Put the current task first. A specific request may need a different format or more detail than your default.
- Check workspace restrictions. If a setting is missing in a managed account, consult your administrator or the current help documentation.
- Retest after editing. Earlier messages are not rewritten when you change the settings.
If a task needs a strict structure or detailed requirements, repeat those requirements in the prompt rather than relying entirely on personalization.
Frequently asked questions
Do Custom Instructions apply to every chat?
They are designed as account-wide guidance. Project instructions are the main exception: they apply inside the project and override global Custom Instructions.
Can I turn them off?
Yes. On the documented web and desktop interface, open Settings → Personalization → Custom Instructions and switch off Enable customization. On mobile, open Settings → Customize ChatGPT. You can also remove the text if you no longer want those instructions used.
Are Custom Instructions the same as a prompt?
No. A prompt describes the task you are asking ChatGPT to perform now. Custom Instructions provide reusable defaults that may shape many tasks. A good prompt still needs to specify the goal, relevant context, constraints, and desired result.
Can they guarantee accurate answers?
No. They can ask ChatGPT to cite sources, flag uncertainty, or avoid unsupported claims, but they cannot guarantee factual accuracy. Verify important information independently, especially for high-impact decisions.