20 ChatGPT custom instructions examples, by role.
By Bernard Huang · Updated
On web or desktop, open Settings → Personalization → Custom Instructions (on iOS/Android: Settings → Customize ChatGPT) and paste one of the sub-500-character role blocks below into 'How would you like ChatGPT to respond?'. Need role rules for just one workspace? Put them in ChatGPT Project Instructions instead so they override global defaults cleanly.
On this page
How to use them.
- Open Settings → Personalization and find custom instructions. Web and desktop can have different controls.
- Choose a short example and edit it. Preserve existing instructions and important uncertainty.
- Review the saved text, then try a factual question, a planning task and a critique request in a new conversation.
- Adjust anything that does not fit. One matching reply does not establish persistent compliance.
OpenAI’s personalization guide explains the layers. The pairings below are editorial suggestions. The generator offers a conservative 1,500-character export budget, not a universal product-limit claim.
The twenty blocks.
Software developer
Pairs with the INTJ or ISTP tuning. 462 characters.
Data analyst
Pairs with the ISTJ or DISC C tuning. 425 characters.
Product manager
Pairs with the ENTJ tuning. 381 characters.
Founder
Pairs with the ENTP tuning. 356 characters.
Marketer
Pairs with the ENFP tuning. 330 characters.
Writer or editor
Pairs with the INFP or ISFP tuning. 347 characters.
Student
Pairs with the any; pair with your type tuning. 312 characters.
Researcher
Pairs with the INTP tuning. 410 characters.
Teacher
Pairs with the ENFJ tuning. 350 characters.
Designer
Pairs with the ISFP or high Openness tuning. 350 characters.
Sales
Pairs with the ESTP or DISC I tuning. 312 characters.
Customer support
Pairs with the ISFJ or DISC S tuning. 333 characters.
Lawyer or contract reviewer
Pairs with the ISTJ tuning. 360 characters.
Finance or analyst
Pairs with the DISC C or Type 6 tuning. 320 characters.
Job seeker
Pairs with the any tuning. 336 characters.
Manager
Pairs with the ENTJ or ESTJ tuning. 324 characters.
Consultant
Pairs with the ENTJ or Type 3 tuning. 310 characters.
Operations or admin
Pairs with the ISTJ or DISC S tuning. 298 characters.
Language learner
Pairs with the any tuning. 320 characters.
Anyone: straight answers
Pairs with the INTJ tuning. 319 characters.
Write preferences you can check.
“Lead with the recommendation” or “ask one question when a missing fact changes the answer” gives you a concrete behavior to evaluate. Broad tone requests can also be useful. We have not established that adjectives fade after a fixed number of turns or that these rules reliably persist in every context.
Combining role and style.
Combine a role template with a library preference only when both fit your work. Delete duplicates and conflicts, then compare responses with and without the combined instructions. A personality label does not validate a role match.
Questions people ask.
What should I write in custom instructions?
Write specific preferences for structure, directness, questions and uncertainty. Pick a role block and edit the lines that do not fit.
What is the character limit?
Use the limit shown in your app. These examples are short; the generator provides a conservative 1,500-character export budget.
Do custom instructions apply to old chats?
OpenAI says updates apply immediately to all chats, including existing chats. They affect subsequent responses; old messages are not rewritten.
Are custom instructions the same as memory?
No. Instructions are explicit preferences you supply. Memory stores context and preferences across conversations; its visibility and controls depend on the product settings. Review saved information in those controls.
Can I combine two blocks?
Yes. Remove conflicting rules and make clear which behavior applies to each task.
Were the role and personality pairings tested?
No. They are editable editorial suggestions. Evaluate the instructions on tasks from your own work.
What changed.
- Clarified preference scope and removed universal field-limit claims; current setup controls vary by client.
- First published, twenty roles.