How do you write custom instructions that actually help?

Custom instructions arrived on July 20 as a beta for ChatGPT Plus subscribers, outside the UK and EU for now. You find them on the web under your name, then Settings, then Beta features. The idea is simple: instead of explaining who you are at the start of every chat, you write it once and ChatGPT takes it into account in each new conversation.

The feature gives you two boxes, each limited to 1,500 characters. Changes apply from your next chat, not the one already open. Those limits shape how to write well for it: you have room for a solid paragraph per box, not an essay, and you will want to test edits in a fresh chat.

What makes an instruction useful

Think of custom instructions as the briefing you would give a capable new assistant on their first day. Four qualities separate helpful ones from the rest.

• Stable. Write what is true for most of your chats: your job, your audience, your usual tasks. "I'm planning a trip to Lisbon" belongs in a single conversation, not in your settings.
• Concrete. "Keep it short" is vague. "Under 200 words unless I ask for more" can be followed and checked.
• Behavioral. Describe what you want ChatGPT to do, not adjectives about it. "Give me three options, then recommend one" works better than "be helpful and decisive."
• Minimal. Every line competes for attention. OpenAI notes that during the beta, ChatGPT may sometimes overlook an instruction or apply it where it does not fit, so a few clear lines tend to hold up better than a long list.

Worked example: a high-school biology teacher

What would you like ChatGPT to know about you?
I teach biology to students aged 15 to 17 at a public high school in Ohio. Many of my students read below grade level, and about a quarter are learning English. I mostly use ChatGPT to plan lessons, write quiz questions and explain concepts in simpler words.

How would you like ChatGPT to respond?
Use plain language at about a grade 8 reading level unless I ask for a different level. When you write quiz questions, include the answer key and mark which learning goal each question checks. Flag any scientific claim you are unsure of instead of guessing. Use US spelling and metric units.

With these in place, a request as short as "Make a five-question quiz on osmosis" should come back at the right reading level, with an answer key, without restating any of it.

Worked example: a small landscaping business

What would you like ChatGPT to know about you?
I co-own a two-person landscaping company serving homeowners in a suburb of Calgary. We do lawn care, planting and small stone patios. I write quotes, customer emails, social posts and the occasional job ad.

How would you like ChatGPT to respond?
Write in a friendly, straightforward tone, like a tradesperson who respects the customer's time. Keep customer emails under 120 words. Use Canadian spelling and prices in Canadian dollars. Never promise dates or prices; leave a [blank] for me to fill in. When I ask for ideas, give at most five.

Note the instruction about blanks. It keeps the model from inventing a price or a start date that a customer might hold you to.

What to leave out

The instructions travel with every conversation, and OpenAI says they are also shared with plugins when you use them; the company's example is a home city passed along to a restaurant-booking plugin. OpenAI may use the text to improve its models unless you turn that off in data controls, though it says it removes personal identifiers first. Its general advice for ChatGPT applies here too: do not share sensitive information.

So leave out your home address, phone number, health details, client names, account numbers and anything about other people that they would not want stored. "A two-person landscaping company in a suburb of Calgary" gives all the useful context; the street address adds nothing.

What should go in each box?

The first box, about you, holds context: your role, your field, your audience, your typical tasks and your level of expertise. Ask yourself what a good assistant would need to know before starting on almost any of your requests. Skills and gaps both help here; "I'm comfortable with spreadsheets but not with code" changes how answers should be pitched.

The second box, about responses, holds rules for the output: length, tone, reading level, format (lists, tables, headings), spelling and units, how many options to offer, and how to handle uncertainty. Write these as instructions you could check against an answer. If you cannot tell whether a reply followed a rule, the rule is too vague.

A quick test sorts any line you are unsure about. If it describes you, it goes in the first box. If it describes the answer, it goes in the second. If it describes only today's task, it goes in the chat.