The analogy
Think of the barber or the seamstress you've gone to for years. The first time you spent ten minutes explaining how you like it: the length, the spot not to touch, the style. Since then you don't repeat it anymore. You sit down, and they already know. That initial explanation has become a permanent preference, valid for every future visit.
Custom instructions are that little preference sheet handed over just once. Instead of re-explaining in every chat that you write for beginners and want short answers, you write it in the settings and the AI "rereads" it at the start of every conversation, behaving accordingly without you opening your mouth.
How it really works
Custom instructions are personal rules that the AI receives at the start of each of your chats, like a layer that adds on top of the company's instructions. In ChatGPT they live in two distinct boxes: one for context about you (role, profession, who your texts are aimed at), one for the way of answering (tone, format, length, language). There's a space cap for each one, around fifteen hundred characters, which is plenty if you're concise. They hold until you change them. Gemini and Claude have equivalent tools, with other names.
They're a different thing from memory. Memory is what the AI notes down on its own by observing you; custom instructions are rules you set deliberately. One you undergo, the other you decide.
What you can do in practice
- In the first box put the useful context: "I'm a lawyer and I write for clients with no legal background" changes all the answers that follow.
- In the second put tone and format. An example ready to paste and adapt:
Answer in English, in a direct and concise way. Get straight to the point, with no preambles like "great question". Use bullet points when you list several things. If you're missing information to answer well, ask me for it instead of guessing.
- Be specific: "professional but simple tone" works better than "write well".
- Update them when your need changes: they're not carved in stone, you rewrite them in thirty seconds.
A common misconception
People think that the more things you write in the instructions, the better the AI will serve you. It's the opposite: the most common mistake isn't writing too little, it's stuffing in a wish list that contradicts itself. "Be brief" and "explain everything in detail" cancel each other out, and the AI ends up in the middle, satisfying neither. Three clear, compatible rules carry more weight than twenty that fight with each other. When you add a rule, check that it doesn't contradict the ones already there.
Frequently asked questions
Are they the same thing as memory?
No. Custom instructions are rules you write and that stay fixed until you change them. Memory is material the AI gathers on its own from conversations. One you set deliberately, the other forms with use. You can use them together.
Do they apply only to me?
Yes, they're tied to your account and apply to your chats. They don't change the AI's behavior for other users. They're your personal layer, on top of the service's rules that are the same for everyone.
Do they work the same way on Gemini or Claude?
The concept is the same: giving permanent instructions about context and style. The name and the menu spot where you find them change, but the idea of a preference sheet that's always active is common. Look for the personalization section in the settings.
Do custom instructions change the model?
No, and believing so makes you expect miracles. They don't retrain anything and don't make the AI more capable: it's always the same model, identical for everyone. What changes is the guidance you give it at the start, like giving different directions to the same driver. They improve how well the answers fit you, not the intelligence of the one writing them.