You keep telling your assistant the same things: "keep it short," "no endless lists," "tell me if you're not sure." With every new conversation, you start from scratch.
Custom instructions fix that. They're permanent rules the assistant reads before every conversation, a bit like a system prompt you write once and for all. ChatGPT offers them, and most other assistants have an equivalent setting, under names and in places that vary by tool and change with updates. Look in your tool's settings or personalization section: the way to write them stays the same everywhere.
Why it's worth ten minutes
A good task prompt says what to do now. Custom instructions say how to work with you, all the time. They save you from repeating basic context, make answers more consistent, and above all fix the habits that annoy you: answers that are too long, a tone that's too enthusiastic, claims without nuance.
It's one of the first things to set up when you start with an assistant, along with privacy and memory. It changes every conversation that follows.
The five sections of a good instruction
1. Who you are
A few lines of useful context: your job or situation, your level on topics you bring up often, what you use the assistant for. "I volunteer for a neighborhood group, I'm a beginner with computers, and I mostly use the assistant to write and get organized." Not your life story: only what changes how it should answer you.
2. What you expect
The kind of help that works for you: step-by-step explanations, honest opinions, several options, questions before answering when your request is vague. That last one is among the most useful: an assistant that asks for clarification avoids an off-target answer.
3. The format
Default length, use of headings and lists, whether to put a summary at the top. "Short answers by default, I'll ask for more detail if needed. No lists for a three-sentence answer."
4. The tone
Casual or formal, plain or technical, and what you don't want: flattery, opening pleasantries, conclusions that repeat everything. Be specific: "be natural" is vague, "no compliments on my questions" is clear.
5. What it shouldn't do
Prohibitions are often the most effective instructions. For example: don't make up figures or references, clearly flag uncertainty, don't present a guess as a fact. That doesn't eliminate errors, but it makes them more visible. To understand why that matters, read why ChatGPT makes things up.
A complete example to adapt
Always answer in English, clearly and directly. I'm [your situation in one sentence] and I'm a beginner at [topic]. By default, keep answers short: I'll ask for more detail if I need it. Use lists only when there really are several items to compare. If my request is vague, ask me a question before answering. If you're not sure about a fact, a figure or a reference, say so explicitly instead of guessing. No compliments on my questions, no closing line that sums up what you just said.
Notice the length: about ten lines. Short instructions beat a novel. The more you add, the more likely they are to contradict each other or be applied unevenly.
Get help writing them
If you don't know where to start, ask the assistant to interview you:
I want to write custom instructions for my AI assistant. Ask me six questions one at a time, waiting for each answer, about: my situation, how I use the assistant, the length and format I prefer, the tone, and what annoys me in its current answers. Then write instructions of under 120 words, in the second person, that I can paste as is. Don't include any sensitive personal information.
Reread the result and cut anything that changes nothing in practice.
What not to put in them
Instructions are read in all your conversations, and they're stored by the service you use. So don't put any sensitive data in them: passwords, account numbers, your exact address, health information, other people's data, your employer's confidential information. Your job or your level is plenty to get better answers.
Also avoid instructions that can't be followed ("never make a mistake") and contradictions ("be very detailed" then "answer in two sentences"). If two instructions clash, the assistant will pick one, and not necessarily the way you'd like.
Test and adjust
Once your instructions are saved, open a new conversation (they usually don't apply to a conversation already in progress) and run three tests:
- A simple question: is the answer short and in the right tone?
- A deliberately vague request: does the assistant ask a question before answering?
- A question whose exact answer you know, on a niche topic: does it flag its doubts?
If a behavior doesn't stick, rephrase that instruction more concretely, or move it to the top. Come back after a week of use: by then you'll know what still bugs you, and you can add or remove a line.
Also remember to check the memory, if your tool has one: it may have kept things about you that add to your instructions. Better to know what's in there.
General instructions or per project?
Your custom instructions apply to everything. But for a specific use (studying a subject, writing for your nonprofit, following a project), some assistants let you create dedicated spaces with their own instructions. In that case, keep the general instructions for what never changes (language, tone, honesty about uncertainty) and put the rest in the dedicated space's instructions. If your tool doesn't allow that, a context block you paste at the start of a conversation works just fine.
To take this further, learn how to create a custom GPT or tailor-made AI assistant dedicated to a single recurring task.
Going further
Writing your permanent instructions is the final challenge of the "Light the Lantern" chapter on Ovrelia: you learn what an LLM is, choose your assistant, set up privacy, instructions and memory, then write a first real request. It's free and plays in your browser, no account needed: play it now.
Instructions set the frame, the prompt does the work: reread how to write a good ChatGPT prompt and pick from our ChatGPT prompt examples. If you're still unsure about the tool, read Claude or ChatGPT: which one to choose. And to test your reflexes, launch the friend challenge: ten timed questions about AI, and a score to beat.
Your instructions are read in every conversation, so keep personal details out of them. To practice spotting what shouldn't go into an AI chat, play the free game Safe to paste?.
Frequently asked questions
What are ChatGPT custom instructions?
They're permanent rules you write once and that the assistant takes into account in your new conversations: who you are, what you expect, the format and tone you prefer, what it shouldn't do. Most other assistants offer an equivalent setting.
Where do I find custom instructions?
They're usually in the assistant's settings or personalization section. The exact name and location vary between tools and change with updates: if in doubt, check your tool's official help pages.
What should I write in my custom instructions?
A few lines about your situation, the kind of help you want, the default length and format, the tone, and prohibitions such as not making up figures and flagging doubts. Short, precise instructions work better than a long text.
Can I put personal information in my instructions?
Only include what helps it answer you, like your job or your level. Avoid any sensitive data: passwords, banking or health information, your exact address, other people's data or confidential information from work.