In ChatGPT
Select your profile picture or initials, open Settings, then Personalization, and look for Custom instructions.
Steps from OpenAI's custom instructions guide, August 2026.
Free guide ยท 02
Setting global instructions
You don't have to live with the factory defaults. Both ChatGPT and Claude give you a place in settings to spell out how you want to be treated. Write it once, then keep refining it. Whatever you put there shapes every conversation that follows.
Five minutes. No technical anything. Works best after you've done a real task or two.
The idea
Stop repeating yourself
Global instructions give you the ability to curate the environment you work in: the kind of output you get, how the tool interacts with you, what it should stop doing.
So if you keep typing the same corrections (shorter, dates first, stop guessing), that correction belongs in your global instructions, not in every conversation. Write it once and it applies everywhere. Then keep reworking it for as long as you use the tool.
One thing these instructions are not: permissions. They shape tone and format, and they cannot let the tool send, spend, or delete anything. Those doors are separate settings, and they stay closed until you open them on purpose.
Step 1
Find where your instructions live
In Claude
Select your initials, open Settings, and look for Instructions for Claude.
Steps from Anthropic's personalization guide, August 2026.
Step 2
Start with one rule
Don't try to write a constitution. The good rules aren't invented in advance; they come out of real conversations. If something bothers you, open settings and make it a rule. These are some of the things that bothered us and the rules we started to develop:
- Don't rush me. The model kept wrapping up while we were still thinking. The rule: I'll tell you when I'm done.
- Dates and action items first. Household answers should lead with what happens when, not with a warm-up paragraph.
- Never make up facts. If the model isn't sure, we want it to say so and say why. "I don't know" beats a confident wrong answer every time.
- Ask instead of assuming. The model kept filling gaps with guesses (ages, dates, budgets we never gave it) and building confident plans on top of them. The rule: if a detail matters and you don't have it, ask me.
The rules aren't ironclad. A banned habit still slips through sometimes, and you remind it the rule exists. They hold most of the time, which is what makes them worth five minutes.
Step 3
Or let it interview you
You don't need to know your rules in advance. Make the tool do the asking, then put the result in your instructions:
Interview me about how I like to work with you, one question at a time: tone, pet peeves, how blunt to be, and what to do when you're not sure about a fact. Five questions max. Then turn my answers into a short set of instructions I can paste into my settings.
Read the draft before you save it, edit anything that doesn't sound like you, and paste it into the setting from step 1. Done.
Going further
Teach it to write like you
Global instructions change how it talks to you. Getting it to write like you takes one more step. Paste in a handful of emails you've sent where the writing is the key part, and ask:
Write a short style guide for my voice: sentence length, how I open and close, how warm I run, words I never use. Prove each rule with a line quoted from my emails. Keep it short enough to live in your instructions.
The quoted-evidence line is what makes this work. Without it, the first pass may come back as flattery: "confident and engaging," which is useless. When every rule has to cite one of your own sentences, the guide gets real, and first drafts start needing a light edit instead of a rewrite.
You are done
Your instructions can keep evolving
Nobody writes their rules in one sitting. Every line should exist because something real bugged you, and the set grows as you figure out how you like to work. Before you close settings:
- Save at least one rule in your instructions, not just said in a chat.
- Read whatever the tool drafted before you save it.
- Remember that instructions shape tone and format, and grant nothing.
- The next time an answer bothers you, turn the fix into a global rule.
This is guide 2 of 3. Next: The family calendar assistant. Back to Getting started with AI.
Find your way in
Join the list and get the free Family AI Starter Kit: practical AI for family life, from three parents who test everything in their own households first.