ChatGPT Custom Instructions, Explained for Beginners

A Standing Setting, Not a One-Off Prompt

Custom Instructions are persistent, user-managed rules that apply automatically to every new conversation you start, without you typing anything special. That's the core difference from a regular prompt: a prompt affects one message, an instruction affects every future chat until you change it. You write it once, in Settings, and it sits quietly in the background from then on, shaping tone and content the same way a job description shapes how a new hire approaches their first week of work.

It's built around two separate text boxes. The first — “What would you like ChatGPT to know about you?” — is context: your role, what you're working on, relevant background. The second — “How would you like ChatGPT to respond?” — is behavior: tone, format, what to avoid, how much detail you want. ChatGPT reads both before generating every response, so the effect is immediate and consistent rather than something you have to keep re-triggering.

Where to Find It

The setup takes under two minutes and applies to every conversation from that point on.

On web or desktop: click your profile icon in the lower-left corner, choose Settings, then Personalization, then Custom Instructions (make sure “Enable customization” is toggled on). On mobile: Settings → Account, then toggle on Custom Instructions. Fill in both boxes, hit Save, and start a new conversation to see the difference — instructions apply to new chats, not retroactively to ones already open.

It's available on every plan, including Free, on web, desktop, iOS, and Android alike — this isn't a paid-tier feature, just one that's easy to miss because nothing in the interface points you toward it.

The Character Limit Tripled

On July 15, 2026, OpenAI raised the limit on each field from 1,500 characters to 5,000, for Plus, Pro, Enterprise, Business, and Education users. That's roughly the difference between a short paragraph and a full page of text — previously, a detailed role description and a full list of formatting preferences often couldn't both fit in the same box, forcing hard trade-offs about what to leave out. At 5,000 characters, most people can fit their role, their working context, and their formatting rules in one place without cutting anything essential.

If you set up Custom Instructions before the update and haven't touched it since, it's worth revisiting — there's now room to be considerably more specific than the old limit allowed, which tends to produce noticeably more consistent results.

The Mistake Almost Everyone Makes

The most common failure isn't writing too little — it's writing a wish list of contradictory demands: “be concise, be detailed, be friendly, be blunt, don't apologize, use examples, keep it short, explain your reasoning.” Instructions that pull in opposite directions cancel each other out, and the output ends up feeling inconsistent because the rules genuinely are.

A few practical guidelines that consistently work better:

• Be specific, not aspirational. “Use bullet points for any list of three or more items” works better than “be well-organized.” A rule you can actually test is a rule the model can actually follow.
• State the negative with a replacement. “Don't use em dashes — use a period or comma instead” gets followed more reliably than “don't use em dashes” alone, which leaves the model guessing what to do instead.
• Keep project-specific rules out of it. If a rule only matters for one recurring task, it belongs in that Project's instructions, not your global Custom Instructions — otherwise it creates noise in every unrelated conversation.

What a Filled-In Set of Instructions Looks Like

Vague instructions produce vague results. A first box like “I'm a small business owner” gives ChatGPT almost nothing to work with. A more useful version: “I run a five-person landscaping business. I handle quotes, scheduling, and customer emails myself. I'm not deeply technical — explain software or automation steps in plain terms, not jargon.” That's specific enough to actually change how ChatGPT responds to a request about, say, drafting an invoice email or troubleshooting a scheduling app.

The second box works the same way. Instead of “be concise,” try: “Default to short paragraphs, not bullet points, unless I ask for a list. Skip the summary at the end unless the answer is genuinely long. Don't hedge with phrases like 'it depends' — give me your best answer and note the exception separately if there is one.” Each line there is a rule ChatGPT can actually check itself against, which is the difference between instructions that shape output and instructions that get quietly ignored.

How This Differs From Memory

They're designed to work together, not compete: Custom Instructions set the baseline rules of engagement, and Memory layers in personal context on top. Worth knowing — OpenAI's own documentation notes that if a saved memory conflicts with something in your Custom Instructions, the memory can quietly win, so it's worth an occasional check of both if a response feels off.

What Custom Instructions Can't Do

Custom Instructions are entirely local to ChatGPT — they don't transfer if you also use Claude, Gemini, or any other AI tool, so each platform needs its own setup. There's also no version history: if you rewrite your instructions, the previous version is simply gone, with no way to revert. And they're not a knowledge base — they shape tone and behavior, but they don't give ChatGPT access to private documents or a custom set of facts the way uploading files to a Project does. For anything that needs actual reference material, Projects is the better tool; Custom Instructions is purely about how ChatGPT talks to you, not what it knows beyond what you tell it directly. It's also worth remembering that long conversations can dilute their influence — the further a chat runs, the more the system prompt's pull can fade against everything that's accumulated since, so if a long-running conversation starts drifting from your instructions, a fresh chat will usually snap it back.

Frequently Asked Questions