What is a system prompt? How it shapes an AI assistant

A system prompt is a block of instructions an AI assistant receives before your conversation starts, telling it how to behave. The app or developer behind the assistant writes it, not you, and the model reads it ahead of every message you send.

What does a system prompt contain?

The system prompt of a support bot for a bicycle shop might read: "You are the support assistant for a bicycle shop. Answer questions about orders, returns, and repairs in under 120 words. If a question is about anything else, say you can only help with the shop. If a customer asks for a refund, hand them to a person."

Real ones run longer, and many apps do not show theirs. Most cover some mix of these:

• Role. Who the assistant is and what it is for, as in the first sentence above.
• Tone. Formal or casual, brief or thorough, and how to address the user.
• Rules. What to stay on topic about, what to decline, and when to hand the person to a human.
• Tools. Which tools it may use, such as search, a calculator, or a calendar, and when to use them.
• Output format. Plain text, a short answer first, a table, or a fixed structure another program can read.

System prompt vs. user message: what is the difference?

Your message is the request: what you want now. The system prompt is standing context: how the assistant should handle every request. The model reads both as text in one conversation, but they have different jobs, and models are usually trained to give system-level instructions more weight when the two conflict. How strongly varies by model and product.

The system prompt also counts against the context window, the amount of text a model can consider at once, so a very long one leaves less space for yours. Our explainer on what a context window is covers that limit.

Why do two apps built on the same AI model behave differently?

Two apps can use the same underlying model and feel nothing alike. One is blunt and good at code. The other is chatty and politely refuses anything off topic. A good share of that gap is the system prompt, along with the tools each app connects, settings such as reply length, and any extra checks the app adds. The model supplies general ability; the system prompt points it at a job.

So testing two assistants with one question can mislead you, because you may be comparing instructions as much as models. How to choose an AI assistant offers a fairer way to compare them.

What can a system prompt not do?

A system prompt is guidance, not a guarantee. The model reads it and usually follows it, but it is not code that enforces rules. Three limits matter:

• It can be ignored or misread. A model may drift from an instruction, especially in a long conversation or when a request is unusual.
• It is not secret. People can often get a model to reveal or paraphrase its hidden instructions, so never put passwords, keys, or private data in one.
• It is not a security boundary. A web page, email, or document the model reads can contain instructions of its own, and the model may treat them as if you had written them. That is called prompt injection, and how prompt injection works shows what it looks like.

Anything that must never happen, such as deleting files or sending money, needs a control outside the model: permissions, limits, and approval from a person. That matters most for assistants that act through tools; see AI agents vs. chatbots.

How do you write your own instructions for an AI assistant?

Many assistants let you save standing instructions. The name varies: custom instructions, personalization, project instructions, or a system message in developer tools. Where they exist, they work like a small system prompt that you write, and the app's own instructions typically still apply too. If your assistant has no such setting, paste the same text at the start of a conversation.

Write down five things:

• Role. What it should act as: "a careful editor of business emails."
• Audience. Who the output is for: "customers who are not technical."
• Constraints. Limits on length, tone, vocabulary, and what to avoid.
• Format. How the answer should look: "the answer first, then at most five bullets."
• When unsure. What to do at a gap: "If a fact is missing, ask me one question. Do not guess."

The last one matters most. An assistant asked for an answer tends to supply one, even when it should say it does not know, which is one reason AI makes things up. A clear fallback instruction can reduce that, though it does not remove it.

For one-off requests, use a checklist for writing better AI prompts. Custom instructions apply to every chat, so keep them to what stays true every time.

Example: vague instructions vs. specific ones

Before: someone saves "Be helpful and write well." The assistant has little to go on. Replies swing between one line and five paragraphs, the tone changes from chat to chat, and it guesses who is reading.

After: the same person writes, "You are an editor of short business emails. The readers are busy customers who are not technical. Keep replies under 150 words, in plain words, with no jargon and no exclamation marks. Start with the answer or the request. Give the draft first, then one line on anything you changed. If you do not know a name, a date, or a price, leave a bracketed blank and ask me instead of inventing it."

Ask each version to reply to a customer whose refund is late. The first may return five cheerful paragraphs that promise a date it invented. The second should return a short draft that opens with the status of the refund and leaves "[date]" blank. Nothing is guaranteed, but when a result is off, you can see which line to change.

Common mistakes when writing AI instructions

• Too long. A wall of rules buries the ones that matter, and every line uses context. Start short and add a rule only when you see the problem it fixes.
• Contradictory. "Be brief" next to "explain everything in detail" forces the model to choose, and you will not know which it picked.
• Vague. "Be professional" means different things to different readers. "Plain words, no jargon, no exclamation marks" is something it can follow and you can check.
• Never testing. After each change, try a few ordinary requests and a few awkward ones before adding more rules.

In a long chat, early instructions can also slip out of view. Keeping a long AI conversation on track covers that.

Where NAMELESS stands

NAMELESS is an AI model that is still in training. It is not yet available, and there is no launch date. It is built on our own architecture, with unlimited context and no need to reread the conversation to continue it. It is also extremely fast, with stronger logic and reasoning, at a cost of pennies compared with today's leading assistants. None of that replaces the advice above: short, specific instructions you have tested help with any assistant. To follow progress, join the NAMELESS waitlist.