Writing a good system prompt
The anatomy of a prompt that makes your chatbot feel purposeful, plus the seven built-in templates.
The system prompt is the chatbot's hidden instructions — the text the LLM reads before every user message. Small changes here produce big differences in how the chatbot behaves.
Where to edit it
Open your chatbot → Edit Settings → the System Prompt card. A 10-row textarea with a character counter below it.
Placeholder text when empty:
You are a helpful assistant for [Company Name]. Your role is to...
The sidebar shows prompt tips and an example. The Use a Template button reveals the seven starting points.
Five ingredients of a strong prompt
- Identity — who the chatbot is
- Scope — what it helps with
- Boundaries — what it refuses to do
- Tone — how it sounds
- Escape hatch — what it does when stuck
The built-in tips panel spells it out:
- Define a clear role: "You are a customer support agent for [Company]"
- Set boundaries: what the bot should and shouldn't discuss
- Specify tone: friendly, professional, casual, formal
- Add escalation rules: when to hand off to a human
- Include key info: business hours, return policy, pricing
The seven built-in templates
Each is a fully-written prompt you can start from. Clicking one replaces your current prompt text.
| Template | Icon | Best for |
|---|---|---|
| Customer Support | 💬 | Answer questions from your docs, escalate to humans when unsure. |
| Sales Assistant | 🎯 | Qualify leads, capture contact info, and book demos. |
| Real Estate Agent | 🏠 | Property listings, schedule viewings, neighborhood info. |
| E-commerce Helper | 🛍️ | Products, shipping, returns, and order status. |
| Healthcare Receptionist | 🏥 | Appointments, insurance, and practitioner info. |
| Restaurant Host | 🍽️ | Menu, reservations, hours, and dietary options. |
| Education Advisor | 🎓 | Courses, enrollment, schedules, and campus info. |
All seven explicitly include contact capture patterns (name, email, phone) where relevant, and include a "don't invent" guardrail.
Iterating from real conversations
The best prompt improvements come from watching real chats. Once a week, open Conversations, skim recent threads, and look for:
- Answers that went off-script → add a line to the prompt about that topic
- Repeated questions the bot dodged → add a Q&A pair (learn more)
- Moments where a human was needed → enable Human handoff (Business plan)
Tips that consistently help
- Write rules, not paragraphs. The LLM follows bullet lists better than prose.
- Name the escape hatch. "If you don't know, offer to connect them with support@acme.com" is a real instruction. "Be helpful" isn't.
- Include the specifics. Business hours, policies, contact channels — put them in the prompt when they matter to every conversation.
- Keep it short. Aim for under 1,000 tokens (~750 words). Long prompts are noise.
Prompt vs knowledge
When the chatbot gives a wrong answer, ask yourself: is it missing facts, or is it using facts the wrong way?
- Missing facts → add training content
- Wrong tone, wrong scope, wrong format → edit the system prompt
Confusing these two is the most common reason chatbots stagnate.