Custom instructions let you state your context and preferences once and stop repeating them. Two things changed since most guides were written: it is now a single Custom Instructions field rather than the old two boxes, and instructions apply immediately to all chats — you no longer have to start a new conversation for them to take effect.
1. Open Settings, then Personalization
Personalization no longer opens in its own modal. Go to Settings → Personalization, where you manage the ChatGPT personality, custom instructions and Memory in one place, then click Custom Instructions. It’s available on all plans, on web, desktop, iOS and Android.
2. Turn on Enable customization
The toggle has to be on before anything you write is applied. This is the first thing to check when instructions appear to be ignored.
3. Write context first, then output rules
You have one field, so give it a shape. Facts about you, then hard rules about the answer:
Context: solo developer building small web apps with Astro and TypeScript,
deployed on Cloudflare. Assume an intermediate level; skip the basics.
Answers: be concise and direct. Give the answer first, then a short why.
Use code blocks for commands. No disclaimers, no "as an AI".
What it does: the top half colors what gets answered, the bottom half fixes how. Rules with a testable shape (“answer first”) survive; adjectives like “be helpful” do not.
You have room to be specific: 1,500 characters on Free and Go, 5,000 on Plus, Pro, Business, Enterprise and Edu. Instructions are not shared with people who open a shared link.
Verify it worked
Ask something you know the model normally answers with a preamble. If the answer still opens with three sentences of throat-clearing, check the toggle is on and cut the instruction down — vague or over-long instructions are the usual cause, not a broken setting. You can edit or delete them at any time and it only affects future conversations.
This is the consumer version of a system prompt: standing text sent ahead of every conversation. For per-project context rather than global preferences, use ChatGPT Projects; on the development side, the equivalent is a CLAUDE.md.
Source: ChatGPT Custom Instructions.