Skip to main content
The system prompt is the single highest-leverage thing you can change about your agent. A good one makes the difference between answers your customers love and answers that make you wince.

​
Anatomy of a good prompt

A working prompt has four parts, in this order:
  1. Identity — who the agent is.
  2. Voice — how it speaks.
  3. Scope — what it’s allowed to talk about.
  4. Rules — what it must never do.

​
Template

Copy this, fill in the bold bits: 
You are **Nova**, the customer support assistant for **Acme Coffee Roasters**.

Voice: friendly, concise, never salesy. Reply in the language the
customer wrote in. Use one short paragraph unless a list is clearly
better.

Scope: you help with orders, shipping, subscriptions, returns, and
our product range. For anything outside that (legal, employment,
partnerships) acknowledge the question and say you'll route it to
the right team, then hand off.

Rules:
- Only answer from the provided knowledge. If the answer isn't in
  there, say you'll check with the team and hand off.
- Never invent prices, SKUs, delivery times, or policies.
- If a customer asks for a human, hand off immediately.
- If a customer is upset, acknowledge first, resolve second.
That’s ~150 words. Don’t go much longer — most of the value is in the first three sentences.

​
What goes where

SectionInclude
IdentityName, company, role (support / sales / onboarding)
VoiceFormal or casual, short or thorough, emoji policy, language rules
ScopeTopics in-scope (and critically, topics out-of-scope)
RulesHard no-goes: no invented policies, no medical advice, hand off for X

​
Common mistakes

  • Giving it facts instead of instructions. “We ship in 3–5 days” belongs in Knowledge, not the prompt. The prompt tells the agent how to behave; knowledge tells it what’s true.
  • Too many rules. After ~10 rules the model starts ignoring some. Prioritise.
  • Ambiguous voice. “Friendly but professional but playful but concise” → confused replies. Pick two.
  • Forgetting the handoff clause. Always say what to do when it doesn’t know.

​
Tuning temperature

Under Advanced, temperature controls creativity:
ValueBehaviour
0.0 – 0.3Deterministic. Same question → almost same answer. Best for support.
0.4 – 0.7Balanced. The default.
0.8 – 1.0Creative. Good for content/brainstorming agents, not support.

​
Iterate with the Try panel

Don’t guess whether your prompt works. Open AI agents → Try and chat with the agent like a customer would. Paste real past questions. When you catch a bad answer, tweak the prompt, try again — 5 minutes of that is worth more than an hour of speculating. See Test an agent.