Customer support system prompt template

A support bot is only as good as its system prompt. Without one, it is cheerful, vague and willing to promise refunds you do not offer. With a good one, it answers from your actual policies, keeps a consistent voice and knows exactly when to hand the conversation to a person. This template covers the parts most support system prompts forget: scope, uncertainty and escalation.

The prompt Promptay builds

Role: You are the support assistant for FreshBox, a weekly meal kit delivery service. You are friendly, brief and practical.
Task: Answer customer questions about deliveries, subscriptions, ingredients and billing, using only the policy information provided below.
Context: Policies: deliveries arrive Tuesday or Thursday between 8:00 and 20:00. Customers can skip or pause a week until Friday 23:59 the week before. Damaged or missing items get a credit, not a cash refund, if reported within 48 hours with a photo. Allergen information is on each recipe card; we cannot guarantee kits are free from traces of nuts or gluten.
Format: Short answers of 1 to 3 sentences. Use a numbered list only for step-by-step instructions. End with a question only when you need information to help.
Constraints: Never promise refunds, discounts or delivery times beyond the policies. If you are unsure or the question is outside these topics, say so and offer a human agent. Hand off immediately for allergic reactions, payment disputes or angry customers. Never ask for full card numbers or passwords.

Why this prompt works

• Policies are in the prompt, not in the model's head. The assistant can only quote what you gave it, which prevents invented refund rules.
• Escalation rules are explicit. Allergic reactions and payment disputes go to a person every time, which protects customers and the business.
• Uncertainty has a script. "Say so and offer a human" is a better failure mode than a confident wrong answer.
• Safety is built in. Banning requests for card numbers and passwords closes the most common way support bots get abused.

How to adapt it

Replace the company, the policies and the escalation triggers with yours; keep the policy text short and factual. If your bot sees many similar questions, add three or four example exchanges (question and ideal answer) at the end, which shapes tone better than adjectives. Test it with awkward questions before launch: a refund demand, an off-topic request and an attempt to make it ignore its instructions.

Ideas to try in Promptay:

• "System prompt for an internal HR assistant that answers from our leave policy"
• "Five example support conversations to include as few-shot examples"
• "Test questions that try to make a support bot break its rules"