To make ChatGPT’s everyday replies clearer, add a few durable preferences in Settings > Personalization > Custom Instructions on the web or desktop app. On mobile, open the menu, choose Settings, then Customize ChatGPT. Tell it how you want answers written, save the changes, and test them with a real question. You can edit the instructions anytime or turn them off. They shape future replies, but they do not guarantee accuracy or override safety rules.
Decide what belongs in custom instructions
Custom instructions work best for preferences you repeat often, not for a one-time assignment. Think of them as a small standing brief: your preferred language, tone, answer length, and the way you want uncertainty handled. They can save you from adding “keep this short” or “use plain English” at the end of every new chat.
Good candidates are stable and specific. For example, you might prefer US English, short checklists, and a clear statement of assumptions. You might also ask for a brief answer first and more detail only when needed. Avoid putting temporary project facts there; those belong in the individual conversation where they are relevant.
Keep personal or confidential information out. Do not include passwords, recovery codes, financial account numbers, health records, client names, or private company details. That is especially important if you later use connectors, shared chats, or third-party tools. A useful preference should describe how you want help, not expose information ChatGPT does not need.
Add your instructions on web or desktop
In ChatGPT on the web or desktop app, open your profile or account menu and choose Settings. Look for Personalization, then open Custom Instructions. The exact labels can change as the product evolves, so OpenAI’s Custom Instructions help article is the best reference if your screen looks different.
Enable the feature if your account shows a switch, then fill in the prompts with a compact description of your preferences. You do not need to use every available character. Short instructions are easier to audit later and leave less room for conflicting requests.
On iPhone or Android, open the side menu, choose Settings, then select Customize ChatGPT. Add the same preferences there and save. OpenAI documents the web, desktop, and mobile paths in its help article; availability and menus can vary by app version or account.
Start with a copyable everyday example
Here is a sensible first draft for someone who wants practical, readable answers:
Write in plain US English. Start with the direct answer, then give a short checklist when useful. Keep routine answers concise. State important assumptions and flag information that needs verification instead of guessing. Ask one clarifying question only when the answer would change materially.
This tells ChatGPT about format and judgment without asking it to pretend it knows facts it does not have. Change the wording to match your real habits. A parent might ask for family-friendly steps; a student might prefer examples before terminology; someone planning work may want a table with next actions and owners. Put only preferences you will actually notice and value.
Do not write instructions that fight each other, such as “always be extremely brief” and “explain every detail.” When preferences conflict, any system will have to choose. Pick one default, then make an exception in the specific chat: “For this one, give the longer version.”
Run a three-question test chat
Saving instructions is not the finish line. Open a new chat and try three ordinary prompts that represent how you use the tool. For example:
- “Give me a simple plan for organizing a busy Saturday.”
- “Explain whether I should back up photos before changing phones.”
- “Help me compare two options when I have not given you enough detail.”
Check whether the replies start directly, use the format you asked for, and identify assumptions rather than hiding them. For the second prompt, compare the response with our Google Photos backup checklist; AI can organize a process, but you should still confirm the current settings on your own device.
If a response is close but not right, revise one phrase rather than rebuilding everything. “Use no more than five bullets” is clearer than “be shorter.” “Say what you are assuming before recommending a purchase” is clearer than “be careful.” Test again in a new chat so you can see the effect on a fresh request.
Use chat-specific directions for exceptions
Custom instructions are a default, not a contract for every task. Put the details that change from job to job in your prompt: the audience, deadline, budget, source material, and desired output. You can say, “For this response, write a detailed email,” even if your default asks for short checklists.
This division keeps your general preferences clean. It also reduces the chance of old project context influencing unrelated questions. When a task needs current facts, ask for sources and check them yourself; custom instructions cannot make a response automatically up to date or correct.
For a different kind of repeatable AI workflow, see our guide to using ChatGPT Voice for an action plan. It shows how a few clear constraints can turn a loose idea into a manageable next step without treating the output as final authority.
Change or turn off instructions when they stop helping
Return to Settings > Personalization > Custom Instructions (or Customize ChatGPT on mobile) whenever your preferences change. Edit the text, save, then try a new chat. If the instructions have become noisy, remove the least useful sentence first. Most people need only a handful of preferences.
To stop using them altogether, turn off the Custom Instructions switch in the same area. You can also clear the fields if you do not want the text saved for later. Changes are intended to affect future conversations; an existing chat may still contain earlier messages and context. Start a new conversation when you want the cleanest test.
If your reason for reviewing settings is account security, that is a separate job. Our guide to checking ChatGPT active sessions explains how to review unfamiliar devices and sign them out.
Keep the expectations realistic
Custom instructions can make the first draft more useful, but they do not turn ChatGPT into a source of truth. It may misunderstand your request, overlook a constraint, or give a plausible but incorrect statement. Read important outputs, verify time-sensitive details with primary sources, and use your own judgment for decisions involving money, health, law, safety, or confidential work.
The best setup is modest: a few durable preferences, one quick test, and an occasional cleanup. That gives you clearer everyday answers with less rewriting while leaving you in charge of the facts and final decision.
Official sources
- OpenAI Help Center: Custom Instructions for ChatGPT
- OpenAI Help Center: ChatGPT release notes