Skip to main content

Prompt guides

System prompts vs user prompts: what each one does

The difference between a system prompt and the messages you type in a conversation — what each is for, when to use which, and how they work together.

All prompt guides · Last reviewed:

What a system prompt is for

A system prompt sets an instruction that applies to the whole conversation rather than one message in it — a persona to hold, a tone to keep, a rule to always follow ("always answer in formal English" or "never suggest a specific dosage"). It is set once, typically before the conversation starts, and a model treats it as standing guidance rather than something to negotiate with each new message.

What a user prompt is for

A user prompt is what you type at each turn of the conversation — the specific question or task for that message. It is where the writing-clear-prompts guidance about context, constraints, format and examples mostly applies, since a user prompt is usually about one concrete thing rather than a standing rule for the whole conversation.

When to put an instruction in the system prompt instead of repeating it

An instruction that should hold for every message — a tone, a persona, a boundary — belongs in the system prompt, so you are not restating it each turn and risking it being dropped or contradicted partway through a long conversation. A one-off request that only applies to the current message belongs in the user prompt. A system prompt is not a security boundary on its own; see what is prompt injection, linked below, for why an instruction in either place can still be overridden by adversarial content elsewhere in a conversation.

Questions people ask

Can a user message override a system prompt?
It depends on how a specific product handles it, and this is not a settled guarantee in general — a system prompt is meant as standing guidance, not an unbreakable rule. See what is prompt injection, linked below, for why treating it as an absolute security boundary is a mistake.
Do I need a system prompt for a simple, one-off question?
No — a system prompt earns its place when an instruction should apply across an entire conversation. For a single question, putting everything in the user prompt is simpler and just as effective.
What kind of instruction belongs in a system prompt rather than a user prompt?
A standing rule that should not need repeating — a persona, a tone, a boundary the model should always respect. A specific, one-time request belongs in the user prompt instead.

Practice on a real conversation

ClawAI’s system prompt setting applies to a whole conversation the same way across every provider it routes to, so a standing instruction does not need to be rewritten per model.