Agent Node — System Prompts That Drive Conversations

The Agent node runs the LLM. The system prompt is the instruction set the LLM follows for the entire time the conversation is in that node. A few paragraphs of plain English about your business, plus a few rules about how to handle calls, is enough for a working agent. This guide covers patterns, common mistakes, and how to write prompts that hold up in real calls.

The anatomy of a good system prompt

A solid system prompt has four parts. Identity: who the agent is, who they work for, what their role is. Knowledge: what the agent knows about the business — services, hours, pricing, policies. Rules: how to handle common situations — booking, escalation, pushback, hang-ups. Style: tone, pacing, what to never say.

Example skeleton:

You are [Agent Name], the AI voice agent for [Business Name], a [industry] in [city].\n\nYou know:\n- Hours: Monday-Friday 9-5, closed weekends\n- Services: tune-ups, brake repair, tire rotation\n- Pricing: tune-up is $45, brake inspection is $60/hour\n\nRules:\n- If the caller wants to book, ask for their name and a callback number\n- If the caller asks for a price we do not have listed, offer to take a message\n- If the caller says 'I want to talk to a human', transfer to (555) 555-1234\n\nStyle:\n- Be warm and patient, never rushed\n- Use plain language at an 8th-grade reading level\n- Never say 'as an AI' or 'I'm just a bot'

Compliance-first prompts (Credit Repair pattern)

For regulated industries, the guardrails live in the system prompt — not as code. The credit repair pattern, in particular, requires lines that say 'NEVER' explicitly. The LLM follows explicit NEVER statements reliably when phrased as imperatives. Example:

You are Alex, an AI assistant with [Company] Credit Services.\n\nYou must NEVER, under any circumstances:\n- Guarantee, promise, or imply a specific credit score increase\n- Claim the company can remove accurate, timely, verifiable negative information\n- Suggest or discuss obtaining a new credit identity, CPN, or EIN-as-SSN\n- Quote a final fee, collect a payment method, or confirm enrollment in-call\n- Ask for or accept a full Social Security Number or full card number by voice\n- Continue past the point where the caller seems confused, distressed, or elderly\n- Give legal, tax, or bankruptcy advice\n\nYou must ALWAYS:\n- Disclose you are an AI and that the call may be recorded at open\n- Confirm you are speaking with the right person before discussing account details\n- Deliver these disclosures before any offer: free self-dispute right, no advance-fee, 3-day cancellation right\n- Immediately transfer to a human if the caller says stop, mentions bankruptcy, raises a legal complaint, or wants to pay/sign\n\nStyle: calm, plain-spoken, patient. Never pressure, rush, or use urgency language.

This is the exact pattern baked into Shopbell's Credit Repair Specialist template (see the template picker under Finance & Legal). The generated workflow uses this structure — you customize the business name and any industry-specific rules.

Context variables you can use in prompts

Your prompt can reference variables that get filled at call time. Common variables: {{caller_phone}}, {{caller_name}} (from caller ID or CRM), {{business_hours}}, {{current_time}}, {{current_day}}, {{workflow_id}}. Enclose them in double braces. If a variable is empty at call time, the agent treats it as blank — the prompt should be written so the conversation still works with empty variables.

Custom variables come from your CRM via the Pre-Call Data Fetch node or HTTP API tool. Example: a real estate agent prompt references {{property_address}} and {{caller_lead_source}} populated from a CRM lookup before the call.

What to leave OUT of the system prompt

Do not put your Twilio credentials, API keys, or any secrets in the system prompt. The prompt is visible in the recording transcript and Recordings history. Do not write long monologues — keep paragraphs short. Do not write 'always be polite' twice — the LLM applies each rule once. Do not list every possible caller question — let the LLM improvise within the rules you set.

Do not put verbal tics in the prompt ('always say awesome!'). The LLM will literally say 'awesome' in every sentence. The prompt sets the LLM's priors; whatever you tell it becomes a habit.

Frequently Asked Questions