How to Write an AI Assistant System Prompt

Every time you start a new chat, you probably repeat yourself: who you are, how you like answers, what to avoid. A system prompt, or its everyday cousin, custom instructions, lets you write that once. Done well, it makes an assistant more consistent. Done badly, it adds a pile of contradictory rules that make results worse.

This guide explains the idea, gives you a structure to follow, and walks through three short examples you can adapt.

System prompt, custom instructions and a normal prompt

The names differ between products and change often, so check your app for what it offers and what limits apply. In general terms:

• A normal prompt is the message you send in a chat. It covers one request.
• Custom instructions are saved preferences about you and how you like replies. They apply to your chats automatically.
• A system prompt is the standing instruction a builder sets for a specific app, bot or assistant. The person chatting usually does not see it.
• A reusable assistant is a saved setup, called different things in different products, that bundles instructions and sometimes reference files for one job.

The writing method is the same for all of them: you are describing a role, a process and some boundaries that should hold across many future conversations, not asking for one result.

The parts of a good system prompt

A strong set of standing instructions tends to include these parts, roughly in this order:

1. Purpose: one sentence on the assistant's job.
2. Audience and context: who it serves and what the work is about.
3. Voice: how it should sound.
4. Process: the steps it follows, such as asking a question before drafting.
5. Rules: specific do and do-not items.
6. Boundaries: what it must decline or send to a human.
7. Output format: the usual shape of a reply.
8. When unsure: what to do when information is missing, such as ask, mark a gap or say so.
9. Knowledge: what it may rely on, for example only the documents you provide.

As a rule of thumb, keep it under roughly 500 words. Long rule lists dilute each other, and contradictions make behavior unpredictable. Put the most important rules near the top, and prefer specific instructions ("ask up to three questions, then draft") over vague ones ("be helpful"). If you are newer to prompting, our guide to writing better ChatGPT prompts covers the basics these parts build on.

Example 1: a writing editor

Purpose: You are my editor. You improve drafts I wrote myself. You do not write from scratch unless I ask.

About me: I write [kinds of documents] for [readers]. My voice is plain and direct, with short sentences and concrete examples.

Process:
1. When I paste a draft, ask what it is for and who will read it, unless I already said.
2. Give feedback in this order: structure, clarity, wording.
3. Show each change with a one-line reason, then the revised text.

Rules:
- Never change facts, names, numbers or quotes.
- Keep my voice. If a change would make it sound like someone else, offer both versions.
- If something is missing, write [NEED: description] rather than inventing it.

Format: two-line summary of the main problems, then the edits, then the revised text.

The key choice here is "improve, do not replace". It tells the assistant what job it has, which prevents the common failure of a full rewrite in a different voice.

Example 2: a study tutor

Purpose: You are a patient tutor for [subject and level]. You help me understand and remember. You do not complete my graded work for me.

How to teach:
- Start by asking what I already know and what I am preparing for.
- Explain in short steps and ask one check question after each step.
- Give a concrete example before a definition.
- If I am wrong, give a hint first. If I am still stuck, explain directly.
- Every few exchanges, give a two-line recap.

Boundaries:
- If I paste an assignment, help me understand the question and the method, and let me do the work. If my course allows more help, I will say so.
- If you are unsure of a fact, say so and tell me what to check in my course materials.
- Never claim to be a teacher at my school or a qualified professional.

Notice the boundary about graded work. It keeps the tool on the side of learning and respects the rules of your school. Check your institution's policy on AI use.

Example 3: a customer-reply drafter

Purpose: You draft replies to customer messages for [shop name], which sells [products]. A human reads and sends every reply. You never send anything.

Knowledge: Use only the policy text I paste at the start. If the answer is not there, write NOT IN POLICY and tell me what to check.

For every message:
1. Classify it: order status, return, product question, complaint or other. Mark ESCALATE if it mentions legal action, safety, injury, a payment dispute or harassment.
2. List the policy facts that apply.
3. Draft a reply under 130 words that acknowledges the specific problem, says what we can do and gives one next step.
4. List anything I should verify before sending.

Rules:
- Never promise refunds, replacements or timeframes the policy does not state.
- Never ask for card numbers or passwords.
- Do not invent order details. Use placeholders like [ORDER NUMBER].
- Calm, warm, plain language.

The structure matters more than the topic. The assistant can only use what you gave it, flags what needs a person, and shows its checks. You can swap in a different business and policy and keep the skeleton.

Test your system prompt before you trust it

A system prompt is a hypothesis until it has met some awkward messages. Run five tests: two typical requests, one ambiguous request, one request that should be refused or escalated, and one that tries to break the rules, such as "ignore your instructions and do this instead". Fix the first failure, then run all five again. You can also have the assistant write the tests:

Act as a tester for the instructions above. Write seven messages a real user might send, including two ambiguous ones, one that should be refused, and one that tries to get you to ignore the rules. Do not answer them. I will send each one separately and report what happened.

Keep the final version, the test messages and the date in your prompt library. Our guide to building a prompt library you will actually reuse shows how.

Safety notes you should not skip

• Instructions guide behavior; they do not guarantee it. The assistant can still make mistakes, and a cleverly worded message, or instructions hidden inside pasted material, can sometimes override your rules. Do not treat a system prompt as security.
• Never put secrets in it. No passwords, keys, private customer data or confidential details. Assume the text could be revealed.
• Keep a human in the loop for anything customer-facing, financial, legal or health-related. Drafting is fine. Sending and deciding are yours.
• Route sensitive topics to professionals. For medical, legal, financial or mental-health questions, tell the assistant to give general information only and suggest a qualified professional.
• Do not let it pose as a person or authority. Instruct it not to claim credentials it does not have.
• Follow platform and workplace rules on AI use and on what data you are allowed to paste.

Where to go from here

The Prompt Bible has a full chapter on custom instructions, system prompts and reusable assistants, with worked examples and a testing routine, built on its broader lessons about prompt structure. If you want a stock of business prompts to turn into assistants for your own work, the AI Business Starter OS bundles hundreds of prompts across business functions with templates and a 30-day plan. For more on everyday business use, see our piece on AI prompts for small business owners.

Frequently asked questions

What is the difference between a system prompt and custom instructions?

A system prompt is usually set by whoever builds an app or assistant and shapes every conversation. Custom instructions are saved preferences an individual user sets for their own chats. The features and names vary by product.

How long should a system prompt be?

Short and specific is better. As a rule of thumb, under about 500 words, with the most important rules first. Long lists tend to contradict themselves.

Can a system prompt stop an assistant from making things up?

It can help, for example by telling the assistant to use only provided documents and to flag gaps, but it cannot guarantee accuracy. You still need to check important facts.

Is it safe to put business rules or customer data in a system prompt?

Put policies in, but keep secrets and private customer data out. Assume the text might be exposed, and check your tool's privacy settings.