Custom instructions
Set your preferences in Settings, choose between instructions, Memory and Projects, and fix rules that don't change the answer.
-
What Should You Put in ChatGPT Custom Instructions? Put in two kinds of text: facts about you that change the answer, and rules for how it should answer. Level 1: Basic promptsFacts about youchange what it says
- Your role and field
- What you know and don't know
- What you mostly use it for
Rules for answerschange how it says it- Length and structure
- When to ask before answering
- When to search and cite
-
Custom Instructions, Memory or Projects: Which Should You Use? Use custom instructions for preferences across all chats, a project for the rules and files of one body of work, memory for things that are useful to recall, and the prompt for anything that matters only right now. Level 2: Context↑ must apply every timeCustom instructionsstable preferences for all chatsProject instructionsrules and files for one body of workMemoryuseful recall, not a guaranteeThe promptdetails for this request onlyspecific to one body of work →
- How Do You Set Up ChatGPT Custom Instructions? Open Settings > Personalization, check that Enable customization is on, write a short list of preferences in the Custom Instructions field, and save it. Level 1: Basic prompts
-
How Do You Improve ChatGPT Custom Instructions? Keep a short list of the corrections you repeat. Level 2: Context
- 1You repeat a correction
- 2Write a checkable rule
- 3Test it in a new chat
No change? Rewrite or delete the rule
Where these settings live now
OpenAI keeps custom instructions under Settings > Personalization on the web, alongside Base style and tone and memory. Its docs say custom instructions are for preferences you want ChatGPT to follow across chats. Codex is separate: OpenAI says personal instructions there are stored in a global AGENTS.md file. So if you use both, you may need to set your rules in two places. OpenAI's Help Center article on this feature refused automated reading on 6 November 2026, so check the exact field names and any character limit on your own settings page.
Treat them as a short list you maintain
People tend to write these once and forget them. A better approach is to start with five lines, then add a rule only after you've given the same correction three times, and remove rules that no longer help. That's a rule of thumb, not OpenAI's advice. Each rule should be something you can check in a reply. This lesson covers how to set them up, what to write, where else context can go, and how to tell whether a rule is working.
What are ChatGPT custom instructions?
They're saved preferences that ChatGPT follows across your chats, such as your preferred response style. You set them under Settings > Personalization on the web, so you don't have to repeat them in every prompt.
What should I put in custom instructions?
Two kinds of text: facts about you that would change an answer, such as your role and what you already know, and rules for how to answer, such as length and format. Leave out details that only apply to one project. Those belong in project instructions.
Why is ChatGPT ignoring my custom instructions?
The rule may be too vague to act on, or another setting may be shaping the reply, such as Base style and tone, memory or project instructions. Rewrite it as something you can check, like a word limit, and test it in a new chat.
Do custom instructions work in Codex?
OpenAI's docs say that in Codex, personal instructions are stored in your global AGENTS.md file, usually at ~/.codex/AGENTS.md. Add your coding rules there, and repositories can add their own AGENTS.md on top.