System Prompt vs User Prompt: What Goes Where (and Why It Matters)
Putting the right instruction in the right place is one of the cheapest ways to make AI output more reliable.
6 min read · Updated
Quick answer
A system prompt sets durable rules, persona, and constraints that apply to an entire conversation — for example, the assistant’s role, tone, output format, and hard limits. A user prompt is the specific request for a single turn. The rule of thumb: anything you want enforced every time goes in the system prompt; anything specific to this one request goes in the user prompt. Mixing them up is a common cause of inconsistent results.
Most chat interfaces and every API expose two kinds of instruction: the system prompt and the user prompt. Understanding the division of labour between them removes a whole class of frustrating, inconsistent results.
What belongs in the system prompt
The system prompt sets the stage. It is read once and shapes every turn that follows, so it is the home for anything durable:
- Role and persona — “You are a meticulous financial analyst.”
- Tone and voice — formal, friendly, terse.
- Hard constraints — “Never give medical advice without recommending a professional.”
- Default output format — “Always answer in markdown with a summary line first.”
- Safety and scope — what the assistant should refuse or redirect.
What belongs in the user prompt
The user prompt is the specific job for this turn: the question, the document to summarise, the data to analyse. It can override or extend the system prompt for one request (“for this answer only, be very brief”), but its job is the task, not the rules.
A worked example
SYSTEM:
You are a support engineer for a SaaS product. Always reply
in three parts: (1) a one-line empathy statement, (2) the
fix as numbered steps, (3) a link to relevant docs. Never
promise refunds. Keep a calm, professional tone.
USER:
A customer says exports are failing with a 500 error after
they upload a CSV larger than 10MB. Draft a reply.Because the format and tone live in the system prompt, every reply across a long conversation stays consistent without you repeating the rules each time.
Model-specific notes
Claude follows system prompts especially closely and works well with tag-structured rules — see best Claude prompts. The principle is the same across GPT-4 and Gemini: persistent behaviour in the system prompt, the task in the user prompt.
Next steps
Once you have a solid system prompt, the rest of your craft is in the per-turn request. Sharpen that with the prompt engineering guide and save your best system prompts as reusable entries in a prompt library.
Frequently asked questions
- What is a system prompt?
- A system prompt is the instruction that sets durable behaviour for an entire conversation: the assistant’s role, tone, output format, and hard constraints. It is sent once and applies to every following turn.
- What is the difference between a system prompt and a user prompt?
- A system prompt sets persistent rules for the whole conversation; a user prompt is the specific request for one turn. Durable constraints belong in the system prompt; the actual task belongs in the user prompt.
- Where should I put output formatting rules?
- If every response should follow the same format, put the formatting rule in the system prompt. If only this one request needs a particular format, put it in the user prompt. Persistent rules go in the system prompt so you do not have to repeat them.
Put it into practice
GeraPrompts is a marketplace of vetted, benchmarked prompts you can run on any model — and where creators keep 85% of every sale.