There's a little box, in every AI tool you use, where you're supposed to tell the model who you are. Most people filled it in once with something vague — "be concise, be friendly" — or left it empty altogether. And the model has been guessing about you ever since.
It's the most underused setting in AI. Custom instructions are the one place a tool lets you hand it durable context on purpose, before the conversation even starts, and almost nobody uses them for more than tone. Fill them in well and every new chat opens already knowing who it's talking to. Leave them thin and you're back to briefing a stranger, every single time.
What custom instructions actually are
Custom instructions are a persistent field where you set context once and the tool applies it to everything you do.
ChatGPT's custom instructions ask who you are and how you want it to respond. Claude does the same through profile preferences and Projects. Gemini keeps it in its saved info and personalization settings. Different label on the box in each tool, same idea underneath: tell the model the things that don't change, so it doesn't have to relearn them every conversation.
Why most people's are weak
Open most people's custom instructions and you'll find a sentence about tone and nothing else. "Keep it short. Be direct." Useful, but it's table manners, not context.
The thing is, telling a model how to talk isn't the same as telling it who it's talking to. Format preferences shape the surface of an answer. Real context — your role, your audience, what you're building, how you write — shapes whether the answer is any good in the first place. Most instructions stop at the surface, which is why the output still feels generic no matter how politely it's worded.
What to actually put in them
Think in three layers, durable to disposable, and keep the durable.
- Who you are. Your role, what you're working on, who you're writing or building for. The facts that are still true next month.
- How you write. The tone and sentence style you want back, the length, the format, and the words you never want it to use.
- The standing context. Your goals this season, your constraints, the people and projects that keep coming up.
Then cut everything that only applies to one conversation. Custom instructions are for what's always true, not for today's task — that's what the chat itself is for.
If you'd rather adapt a real one than start from a blank box, here are copy-paste examples for founders, writers, freelancers, developers, and marketers.
How to write them, tool by tool
The fields differ a little, but the move is the same in all three.
- Say who you are — role, work, audience.
- Say how you write — tone, length, format, and your never-words.
- Add the durable context — the standing facts the model should always assume.
- Cut the one-offs — anything that matters for only a single chat.
- Write it once, paste it everywhere — draft it as one document, then drop the relevant parts into ChatGPT, Claude, and Gemini.
That last step is where most of the pain hides, so it's worth pausing on.
The catch: you're filling in three boxes, not one
Here's what nobody tells you. Fill in ChatGPT's custom instructions beautifully, and Claude still knows nothing. Set up Claude, and Gemini's a stranger. Each tool keeps its instructions locked inside itself, so "set it once" quietly becomes "set it once per tool, forever, including tools that don't exist yet."
That's the same wall every method hits, and it's why it helps to see how the options actually compare. The fix isn't a better box. It's keeping your context somewhere that isn't trapped in a box at all.
Where RUMO fits
Write your context once as a context anchor — a plain document you own — and pasting it into each tool's custom-instructions field becomes a thirty-second job instead of a rewrite. The context lives with you; the tools just borrow it.
And if a single field feels too small for everything that matters, the six context anchors take the same idea further — durable pieces of you, written once and reused across every tool you open.
Start with one
That empty little box was never really the problem. The problem was having nothing durable to put in it.
So write the durable part once. Start free with your Personal Constitution — about thirty minutes — and you'll have real context to drop into ChatGPT, Claude, and Gemini, instead of one more vague line about being concise.
Frequently Asked Questions
- What are custom instructions?
- Custom instructions are a setting in most AI tools where you tell the model, once, who you are and how you want it to respond — so it carries that context into every new chat instead of starting blank each time. ChatGPT, Claude, and Gemini each have their own version of this persistent context field.
- What should I put in ChatGPT's custom instructions?
- Three things: who you are (your role, what you're working on, who you're writing for), how you want it to respond (tone, length, format, the words to avoid), and the durable context it should always assume. Skip the one-off preferences and focus on the facts about you that stay true across most of your conversations.
- Are custom instructions the same across ChatGPT, Claude, and Gemini?
- No. Each tool has its own version — ChatGPT's custom instructions, Claude's projects and profile preferences, Gemini's saved info and personalization — and they don't share. Fill one in and the others still know nothing about you, so you end up writing your context three separate times.
- How do I write good custom instructions?
- Lead with the durable facts: your role, your audience, how you write, and the response style you want. Be specific and concrete, cut anything that only applies to a single conversation, and keep it tight enough that the model can actually use it. Write it as a plain document first, then paste the relevant parts into each tool.
- Do custom instructions work across different AI tools?
- Not on their own — each tool's instructions stay locked inside that tool. To avoid rewriting them everywhere, keep your context as a portable document (a context anchor) you own, then paste it into each tool's custom-instructions field. You write it once and reuse it across ChatGPT, Claude, Gemini, and any agent.
