Writing a system prompt
The system prompt is the standing instruction that defines your bot's role, tone and rules. Promptly sends it with every message, so it shapes every reply. Edit it in
Dashboard → Settings🖼️ [Image] — The System Prompt textarea with the cost-reminder info banner below it.
The textarea shows the placeholder
You are a helpful assistant for [Company Name]...What a good prompt covers
- Role & company — who the bot is and who it works for.
- Tone — formal, friendly, concise, etc.
- Scope rule — tell it to answer only from your knowledge base and product catalog, and to avoid guessing. The bot's reasoning is already bounded to your data (see How retrieval works), but a clear scope rule reinforces it.
- Escalation cue — when to hand off to a human, if you use Escalation.
You don't need to add language instructions — the bot automatically detects and replies in the visitor's language. See Languages & cost tracking.
Example
textYou are Ava, the support assistant for Acme Tools. Be warm, concise, and professional. Only answer using the company knowledge base and product catalog. If you don't have the information, do not guess — say you're not sure and offer to connect the visitor with a human agent. Never discuss pricing for products that are out of stock.
The fallback message
The Fallback Message section (placeholder
I'm not sure about that. Please contact our support...Token-cost tradeoff
Because the system prompt is sent on every request, a long prompt is paid for on every message. A 400-word prompt adds its token cost to thousands of conversations. Keep it tight: state the rules once, drop examples the model doesn't need, and let your knowledge base carry the facts. See Languages & cost tracking.