System Prompt
"System Prompt is the standing instruction block a language model reads before every turn, setting the agent's role, rules, tone and handoff conditions."
What is a System Prompt?
System Prompt is the standing instruction block a language model reads before every turn, setting the agent's role, rules, tone and handoff conditions. Related product: Chat Agents.
Reviewed by Atul Diwedi, Co-Founder · Last reviewed
Deep Dive
The system prompt is the agent's standing orders. It is read before every customer message, it outranks the conversation in the model's attention, and it is the single place where a business says who the agent is, what it may do, what it must never do and when it should hand the conversation to a person. In a voice or chat agent it sits above the retrieved passages, the tool descriptions and the conversation history, which are assembled around it at runtime.
How it works. The platform builds a prompt for each turn: the system prompt first, then the tool definitions, then the retrieved passages for the current question, then the conversation so far, then the latest message. The model produces a reply or a tool call conditioned on all of it, with the system prompt acting as the constant. A working system prompt for a receptionist agent has a recognizable structure: identity and scope (one paragraph), hard rules (a short list: no medical advice, no price quotes outside the price list, confirm name and number before booking), tone and length rules, transfer conditions, and a fallback instruction for questions outside scope. Everything factual lives elsewhere, in documents, so the prompt stays short and stable.
Why it matters for risk and brand. Every embarrassing agent screenshot traces to a system prompt that either did not cover the case or covered it badly. The prompt is where 'never promise a delivery date' and 'always offer a person after two frustrated messages' are enforced. It is also the brand: the difference between an agent that sounds like your best front-desk person and one that sounds like a generic assistant is a few lines of tone guidance and examples.
Example. A spa's chat agent system prompt reads, in part: 'You are the front desk for Marigold Spa. Answer booking, pricing and service questions from the provided documents only. Never give medical advice; for questions about skin conditions or pregnancy, say a therapist will call back and collect a phone number. Keep replies to three sentences. If a customer asks for a person or sends two annoyed messages, offer a handoff.' A customer asks whether a particular facial is safe during pregnancy. The agent does not answer from the service description; it follows the rule, offers a therapist callback and collects the number.
System prompt vs AI guardrails. The system prompt is one layer of guardrails, and the one the model can disobey under pressure. Harder guardrails sit outside the model: output filters that block a reply containing a price not in the price list, tool permissions that stop the agent from issuing refunds, and routing rules that force a transfer on certain keywords. Put the rule in the prompt for behavior, and in a guardrail when the cost of the model ignoring it is unacceptable.
Common pitfalls.
- Pasting the entire FAQ into the prompt instead of into the knowledge base.
- Vague tone adjectives ('professional, friendly, empathetic') with no example of what that sounds like.
- No instruction for the out-of-scope case, so the model improvises.
- Editing the live prompt without a version and a test run.
Botcadence agents hand off to a person with the full transcript when the prompt's transfer conditions are met; see human handoff and chat agents.
Related Technology
See System Prompt in a Botcadence agent
Start on the Free plan and configure an agent that uses System Prompt on your own website chat or WhatsApp number. No card needed.
Get Started FreeWebsite chat live in about 5 minutes
Common Questions
System Prompt FAQs
Identity and scope, a short list of hard rules, what to confirm before taking an action, reply length, transfer conditions, and what to do when a question is outside the documents. Keep facts out; put them in the knowledge base.
Assume yes. Models can be coaxed into repeating instructions, so put no secrets, internal pricing floors or staff contact details in it. Write it as if it were public, and keep sensitive logic in tools and permissions outside the model.
Share the rules and scope, but keep separate tone and length sections. Voice needs shorter sentences, spoken number formats and confirmation read-backs; chat can use links and lists. One prompt with channel-specific sections is a workable compromise.
Run a test set that targets the rule directly and count violations. If a rule is skipped, move it earlier, make it more concrete, or add a one-line example; if it is still skipped, enforce it outside the model with a filter or tool permission.