AI configuration - system prompt, persona and temperature
What AI configuration controls
Open Settings › AI Configuration (shown as Agent in newer dashboards). It has five parts:
| Section | What it sets |
|---|---|
| AI Persona | A generated description of your business that seeds the prompt. You can regenerate it. |
| Name and greeting | The agent's display name and the first message visitors see. |
| System Prompt | Your instructions to the agent. Can be reset to the default. |
| Temperature | How varied the wording is, from 0 to 1. Default 0.3. |
| Behavior | Show sources, proactive lead capture, require email before chat. |
Other settings that shape answers live elsewhere: the reply language under Settings › General, the chat model in the model selector, and how strictly the agent sticks to your content in hallucination prevention.
System prompt
The system prompt is the instruction set the agent reads before every reply. It is the strongest lever you have over tone and boundaries.
Use it to:
- Set the voice. Formal or relaxed, short or detailed.
- Give the agent a role. "You are the support agent for Acme Outdoor."
- Draw lines. "Never promise a refund outside the 30-day policy."
- Say when to hand off. "Offer a person when someone mentions a legal complaint."
Whatever you write, the agent still answers from your indexed content and says it does not know when that content is silent.
System prompt patterns
Three starting points. Replace the company name with your own; the prompt has no template variables.
Pattern 1: support agent.
You are a knowledgeable customer support agent for Acme Outdoor. Answer questions from the knowledge base. Be concise (under 100 words). Cite the source when relevant. If the user is frustrated or asks for a human, suggest they say "talk to a human". Don't promise refunds or discounts beyond stated policy.
Pattern 2: sales engineer.
You are a sales engineer for Acme Analytics. Help prospects understand our product, pricing, and integrations. Be enthusiastic but accurate. If asked about pricing for Enterprise, suggest booking a demo. Don't reveal internal roadmap details.
Pattern 3: internal helpdesk.
You are an HR helpdesk for Acme Corp. Answer employee questions about benefits, PTO, payroll. Be empathetic and clear. Cite policy documents. For sensitive issues (harassment, legal), suggest contacting HR directly via email.
Adapt to your context. Tone tweaks compound across thousands of interactions.
Temperature
Temperature controls how much the wording varies between replies. The default, 0.3, keeps answers steady and close to your content, which suits support. Raise it a little if replies read as stiff; lower it toward 0 for the most predictable phrasing. The slider runs from 0 to 1.
Behavior switches
- Show sources in responses. Shows the documents behind each answer on every channel. On by default.
- Proactive lead capture. Lets the agent ask interested visitors for their contact details. See collect_lead.
- Require email before chat. Visitors enter an email before they can start a conversation.
Drafts and publishing
You can work on the prompt and temperature without touching live channels:
- Save Draft stores your changes. The chat playground answers with the draft.
- Test real questions in the playground.
- Publish pushes the draft to the widget and every connected channel. Discard throws it away.
In newer dashboards, saving applies straight away with an Undo option, and a leftover draft is offered for publish or discard rather than mixed into the form.
Common pitfalls
The agent ignores part of the prompt. Long prompts dilute each rule. Put the rules that matter most first and keep them short.
Answers got worse after an edit. Reset the prompt to the default, then add your changes back one at a time, testing each in the playground.
Published but nothing changed. Check you pressed Publish, not only Save Draft.
FAQ
Will changing the system prompt affect ongoing conversations?
Yes. The published prompt is read on every reply, so the next answer in any open conversation uses it.
Can I write the prompt in another language?
Yes. The reply language is set separately under Settings › General, either matching the visitor or fixed to one language.
How long should my system prompt be?
Long enough to cover your voice and your hard rules. If you are past a page, you are probably restating what your content already says.
Can I test prompts without deploying?
Yes. Save a draft and use the chat playground.
Does the agent follow my prompt strictly?
Mostly, but models can drift. For anything that must never happen, rely on settings and skill limits, such as audience tags, rather than prompt wording alone.