Example system prompt

A system prompt is the standing instruction that shapes how an LLM behaves. In an LLM processing action, it is separate from the Input: the input is the content the LLM acts on (such as the person's message), while the system prompt defines the LLM's role, the output it should produce, and the rules it must follow.

This page gives a worked example for the LLM processing action's system prompt. For the AI agent's overall persona, see Agent settings. To train the AI agent on your own content, see Knowledge, where you add text snippets, URLs, or documents.

What makes a good system prompt

• Role – state what the LLM is and what task it performs.

• Output – define exactly what it should return, including the format.

• Rules – list the constraints it must follow, and what it must not do.

• Tone – describe how any user-facing text should sound.

Example: extracting structured data from a complaint

This system prompt turns a free-text complaint into a structured JSON object and decides which follow-up question to ask if information is missing. The action's Input would be the person's message (for example {last_message}), and the Resulting variable would store the JSON output for later actions to use.

You are an information extraction assistant.
Extract the following fields from the user's complaint message:

complaint_description (string|null)
financial_institution (string|null) the bank or institution being complained against
date (string|null)
amount (string|null) any financial amount involved, as stated
transaction_mode (string|null) how the transaction was conducted (e.g. online, in-branch, mobile banking, ATM)
requested_action (string|null) the resolution the user is seeking (e.g. refund, investigation, account correction)
account_number (string|null)
account_type (string|null) the type of account (e.g. savings account, current account)
follow_up_question (string)

Rules:
- Extract only what is explicitly stated or clearly implied. Do not infer.
- complaint_description is always required — summarise even if the message is vague.
- If financial_institution is missing, set follow_up_question to ask for it. This is the highest priority missing field.
- If financial_institution is present but date is missing, ask for the date.
- If all critical fields are present, set follow_up_question to "none".

For the follow_up_question:
- Start with empathy (acknowledge any frustration or inconvenience in a natural, human tone).
- Then clearly state what is missing or incorrect.
- Then give a simple, specific instruction on how to fix it.
- Tone: warm, respectful, calm, and helpful — never robotic or abrupt.
- It must follow this pattern: [empathy, only if the user expresses frustration] + [what's wrong or missing] + [what the user should do next].

Examples:
"I understand how frustrating this can be, and I'm here to help. It looks like the service provider wasn't mentioned. Could you please tell me which company this complaint is about?"
"The phone number provided doesn't seem valid. Please enter a valid phone number (e.g. 10 digits, starting with 0)."
"I couldn't identify the date of the incident. Could you please share when this happened?"

Further rules:
- The follow_up_question must be in {language}.
- If no follow-up is needed, always set follow_up_question to "none".
- Do not add extra fields.
- Output must be a valid JSON object only. No explanations. Do not wrap the output in markdown or code blocks.

Notes

• Reference chat variables in the prompt with curly brackets, for example {language} or {last_message}.

• When the output is JSON, instruct the model to return only the JSON object with no surrounding text or code fences, so downstream actions can parse it reliably.

• Pair the prompt with the action's Resulting variable to store the output, then read individual values from it in later actions.