Key takeaways
- OpenAI lists custom instructions for all ChatGPT plans on web, iOS and Android, so free accounts can use them.
- OpenAI's help article sets a 1,500 character limit per field, and changes apply only to new chats, not conversations already open.
- Specific, checkable rules such as "lead with the answer" or "under 150 words" work far better than vague requests like "be professional".
In this story 5 sections
You keep typing the same thing at the start of every chat: keep it short, skip the pep talk, I work in sales, use US spelling. ChatGPT custom instructions exist so you can stop doing that. You write your preferences once in settings, and ChatGPT applies them to every new conversation. This guide is for anyone using ChatGPT on the web or in the phone app, on a free or paid plan, and setup takes about ten minutes. Most of that time goes into writing instructions that actually change the answers you get, which is where the examples below come in.
Before you start writing your instructions
A little prep saves you from rewriting everything a week later.
- Know your plan and device. OpenAI's help article on custom instructions lists the feature for all plans on web, iOS and Android, so a free account works.
- Collect your repeat requests. Scroll through recent chats and note what you asked for more than once ("shorter", "give me a table", "no jargon").
- Decide what ChatGPT should know about you. Your job, your skill level in the topics you ask about, and anything that changes the right answer, like living in the US or working on a Mac.
- Leave out anything private. Skip account numbers, passwords, health details and anything you would not paste into a work email. OpenAI's help article notes that custom instructions may be used to improve its models.
- Mind the length. OpenAI's help article sets a 1,500 character limit on each response field, which is roughly 250 words per box.
Opening the custom instructions screen
OpenAI has moved these settings a few times as ChatGPT has added personalization features, so look for the word Personalization rather than an exact old menu.
On the web or the desktop app
- Sign in at chatgpt.com.
- Click your profile name or picture in the corner of the sidebar.
- Choose Settings, then open Personalization. (OpenAI Academy's guide describes the same route as going to your profile and clicking Personalize.)
- Find the custom instructions fields and the "about you" section.
On iPhone or Android
- Open the ChatGPT app and tap your profile or the menu to reach Settings.
- Look for Personalization and the custom instructions option. Older versions of the app placed it under Account, so check there if your app has not updated.
- Make sure the feature is switched on. In the app, a toggle turns custom instructions on or off without deleting what you wrote.
If you use ChatGPT on your phone a lot, our guide on how to use ChatGPT on iPhone covers the rest of the app's settings.
Writing the two parts: about you and how to respond
Custom instructions have two jobs. One part tells ChatGPT who you are. The other tells it how you want answers shaped. Keep them separate, because mixing them makes both harder to follow.
Step 1: describe yourself in a few useful facts
Only include details that change an answer. "I like hiking" rarely matters. "I'm a high school chemistry teacher in Ohio" changes examples, reading level and which state rules apply.
Example that works:
"I run a two person bakery in Austin. I handle marketing, ordering and payroll myself. I'm comfortable with Google Sheets but not with formulas beyond SUM. Prices and laws should be US based."
Step 2: tell it how to answer
Be concrete. "Be concise" is vague. "Lead with the answer in one or two sentences, then details only if needed" is something the model can actually follow.
Example that works:
"Give the direct answer first. Use plain English and short paragraphs. Use bullet lists for steps. Skip compliments, motivational lines and closing summaries. If you're unsure or a fact may be out of date, say so. Ask one clarifying question when my request is ambiguous instead of guessing."
Step 3: add examples for the way you work
Here are a few starting points you can paste and edit.
- For writing help: "Match my tone when editing. Keep my wording where it's fine and only fix what's unclear. Show changes as a short list after the revised text."
- For learning a topic: "I'm a beginner. Explain new terms the first time you use them and give one everyday comparison."
- For coding: "I use Python 3 on a Mac. Give complete runnable code, then a two line explanation. Don't rewrite parts of my code I didn't ask about."
- For job searching: "I'm applying for project manager roles in healthcare. Keep cover letters under 250 words and avoid generic phrases."
For more on phrasing requests, see our guide to writing better AI prompts. The same habits that improve a single prompt make your custom instructions stronger.
Step 4: save and start a new chat
- Click or tap Save.
- Open a new chat. OpenAI's help article says the settings apply to new chats, and changes show up only in future conversations, not ones already open.
- Ask something you ask often and compare the answer with an older one.
Testing and tuning your custom instructions
Your first draft will rarely be perfect. Spend a few days using it, then adjust.
- Run three normal requests. Pick tasks you actually do, such as an email, a quick explanation and a list.
- Check each rule. Did it lead with the answer? Did it drop the pep talk? Did it use US units?
- Fix the rule, not the chat. If one rule is ignored, rewrite it to be more specific or move it higher in the box. Instructions near the top tend to be easier to spot when you reread them, and clear wording matters more than length.
- Remove rules that fight each other. "Be extremely brief" and "always explain your reasoning in detail" pull in opposite directions.
How instructions fit with memory, personality and projects
Custom instructions are not the only way ChatGPT adapts to you, and knowing the difference helps you put each preference in the right place.
- Memory lets ChatGPT carry details forward from past chats on its own. Our explainer on ChatGPT memory shows how to view and delete what it saved. Use custom instructions for rules you always want applied, and let memory handle smaller facts.
- Base style and tone sits under Settings, then Personalization, according to OpenAI's ChatGPT personalization docs. It changes how ChatGPT sounds, not what it can do.
- Project instructions apply only inside one project. OpenAI Academy suggests using them to set the style for a specific body of work, so a client project can follow different rules than your everyday chats.
When your custom instructions don't seem to work
If answers look the same as before, work through these checks.
- You're in an old chat. Changes apply to new conversations only, so open a fresh one.
- The toggle is off. On mobile, confirm the feature is switched on in settings.
- A project is overriding you. Inside a project, its own instructions shape the chat. Test outside the project.
- The rule is too vague. Replace words like "better" or "professional" with something you could check, such as "under 150 words" or "no exclamation points".
- You hit the limit. If text won't save, trim it. Cut examples first, keep the rules.
- You want them gone. OpenAI's custom instructions help article says to clear both fields and save on the web, or switch the toggle off in the app. Past chats keep their old behavior unless you delete them.
One more point if you share chats: OpenAI says your custom instructions are not shown to people who open a shared chat link, so you don't need to clean them up before sharing.
If you're still deciding whether a paid plan is worth it for these features, our comparison of ChatGPT Plus vs Free breaks down what changes. You can find more guides like this in our AI section. A good final check: reread your instructions a month from now and delete anything you no longer need, because shorter instructions are easier for you to maintain and easier for ChatGPT to follow.
Sources
Share this story
Spotted an error? Tell us


