Rules for chat: ChatGPT, Claude, Gemini, Grok

🧭 Rules, projects and memory

The three things that get mixed up most often in the settings.

Rules

You write them. Every AI calls them something slightly different: custom instructions, instructions, personalization.
The AI gets them at the start of every new chat and they apply until you change them.
This is where you put who you are, how it should answer you and what bothers you about its answers.

Project

A folder of chats on one topic, with its own instructions and files.
Handy when you need a different style for one thing, say work and music.
Project instructions only apply inside it. In ChatGPT they take priority over the global ones.

Memory

What the AI saves by itself from your conversations. It may not be accurate.
Go through it in the settings now and then and delete what's wrong or what you don't want there.
Write what matters to you into the rules. Don't rely on memory.

📍 Where to put the rules

For all chats, or just for one topic. The names in the settings change now and then. If you can't find the item, look for the words Personalization or Instructions.

💬 ChatGPT

For all chatsSettings → Personalization → turn on Enable customization → the Custom Instructions field

On mobile: Settings → Customize ChatGPT.
For one topic: Projects → your project → project instructions. They take priority over the global ones.
On paid plans the limit is 5,000 characters. If the text doesn't fit, cut the least important parts.

🟠 Claude

For all chatsYour initials at the bottom left → Settings → Your Claude → Instructions

In the older layout: Account → Instructions for Claude.
For one topic: Projects → your project → Project instructions. Paid plans only.
It handles longer text, but here too the AI follows shorter rules better.

🔷 Gemini

For all chatsSettings & help → Personal context → Your instructions for Gemini → Add +

In some places the section is called Personal Intelligence. All users have it.
For one topic: on a computer, in the side panel Gems → New Gem.
From 17 November 2026 Google will start converting Gems on personal accounts into Skills, and this path will change.

⚫ Grok

For all chatsProfile at the bottom left → Settings → Personalization → Custom Instructions

On mobile, look for Customize Grok.
For one topic: Projects → your project → Instructions.
Keep the text under 8,000 characters. I couldn't find an official guide from xAI; the path comes from other people's guides and may change.

Another AI, like Microsoft Copilot, DeepSeek, Mistral or Perplexity: look in the settings for custom instructions. If it has none, paste the rules at the start of the first message in a new chat.

✍️ How to write rules the AI follows

Based on the guides from Anthropic, OpenAI and Google for their latest models. The same standards apply to the whole site; more in How to write prompts.

What you want and why

“Keep answers short, I read them on my phone” works better than “Don't write long answers”.
The reason helps the AI even in a situation the rule didn't think of.
Instead of a prohibition, write what you want instead.

No shouting

New models take capital letters and words like NEVER, FORBIDDEN or MUST too literally and then overdo the rule.
A plain sentence with a reason works more reliably.

Allow “I don't know”

The rule “when you're not sure about something, say so” cuts down on made-up answers more than “answer with 100% certainty”.
For facts and numbers, ask for a source and a date so you can check them.
Ask the AI to separate what it has verified, what it's only inferring and what it doesn't know. Then you know what to trust.

Be specific

The AI reliably sticks to language, length, format and form of address. “Be professional” means something different to every model.
Describe the role by what it helps you with and for whom. “You are an elite expert” doesn't improve the result.

Keep it tidy

Headings by topic and one rule per line. Then the rules are easy to edit.
Shorter is better. Every extra rule takes attention away from the others.

Test it

After saving, open a new chat and ask something ordinary.
When the AI doesn't follow a rule, rewrite it more specifically, for example with an example. Exclamation marks don't help.

🔁 One set of rules for all four

When you use several AIs, you don't have to write the rules again for each one.

How to do it

Write plain text with headings, like in the template below. All four understand it.
Stay under 5,000 characters, which is ChatGPT's limit. Then you can paste the same text everywhere without changes.
Put things only one AI can do into a separate section at the end. Elsewhere you just delete it.

Where to keep the original

Save the rules in your notes or in a file. AI settings are not a backup.
When you change something, edit the original and paste it everywhere again.
Add the date of the last edit so you know which version is where.

When it doesn't work

A change reliably takes effect only in a new chat.
In a very long chat the rules from the start get lost. Start a new one.
When the rules and the project instructions contradict each other, the AI follows only one of them. Check both.

Don't want to write from scratch? The rules generator asks you 15 things and makes a version for each AI separately. Everything is computed in your browser only.

📋 A template and prompts to copy

Replace the yellow parts in square brackets with your own. Delete whatever doesn't apply to you.

Rules template ChatGPTClaudeGeminiGrok

# My rules for AI

## About me
- My name is [name].
- Write in English. When you use a technical term for the first time, add one sentence on what it means.
- [Who you are and what you do, one sentence.]
- My level in [field]: [beginner / intermediate / pro].

## How to answer me
- Start straight with the answer. Skip the intro, repeating my question and apologies, they slow me down.
- A simple question gets a short answer. When it's a decision or a complex problem, go into depth.
- When there are several options, recommend one and say why. A bare list doesn't help me.
- When you're not sure about something, write “I don't know” or what you're basing it on. Admitted uncertainty is more useful to me than a made-up answer.
- For important claims, separate what you've verified from a source, what you're only inferring and what you don't know.
- For facts, numbers and prices, write where they come from and as of what date, so I can check them.
- When something doesn't work, write why and suggest how to fix it.
- Don't shorten or rewrite my texts until we agree on it. When you see something that could be better, suggest it.

## Format
- Paragraphs and bullet points. A table only for comparisons.
- Code always in a block with the right language.

## When you're missing something
- When you're missing important information, ask before you start.
- For small things, pick a sensible option and say which one.

## My own rules
- [What annoys you most about AI answers and what you want instead.]

## Only for [ChatGPT / Claude / Gemini / Grok]
- [Things only this AI can do. Delete this section in the others.]

Help me write the rules ChatGPTClaudeGeminiGrok

The AI interviews you and writes the rules for you.

I want to write my own rules (custom instructions) that you'll follow when answering me in all chats.

Ask me 6–8 questions one by one. Always just one, and wait for the answer. Find out who I am, what I use you for, how long I want answers to be, what tone suits me and what bothers me about AI answers.

Then write the rules from my answers:
- in English, as bullet points under the headings About me, How to answer me, Format and My own rules,
- each rule as what I want, and for the important ones briefly why,
- no capital letters and no words like NEVER or MUST,
- all of it under [5,000] characters, so it fits into ChatGPT, Claude, Gemini and Grok.
Don't make up anything that doesn't follow from my answers.

Check my rules ChatGPTClaudeGeminiGrok

For rules you already have that the AI still doesn't follow.

Here are my rules for AI:

<rules>
[paste your rules]
</rules>

Go through them and find:
1. rules that contradict each other,
2. rules written as shouting (capital letters, NEVER, MUST) or only as a prohibition without what I want instead,
3. vague rules where you can't tell whether the AI followed them (like “be professional”),
4. things only one specific AI can do.

For each finding, write why it's a problem and suggest better wording. At the end, give the whole corrected version under [5,000] characters. Leave rules that are fine unchanged.

Next: Rules for agents → ← Back to Set up AI