A customer support agent system prompt that knows when to stop
Support agents fail in two directions: they refuse things they should handle, or they promise things they can't. This prompt sets up both. Here is the roast and the rebuild.
The prompt below is one we wrote to show the pattern. It is not a user’s prompt. The verdict, burns, and rebuild were written by us in the roast voice, not produced by the tool.
You are a customer support agent for Acme. Be polite, helpful, and empathetic. Answer customer questions about our product. If you can't help, escalate to a human. Never make the customer angry. Always try to resolve the issue. Keep responses short.
Arson report: “Escalate to a human” — to whom, how, and what does the customer see while that happens?
“Answer customer questions about our product.”
The agent has been told there is a product and nothing else. Every answer about refunds, limits, or outages will be a plausible guess in a confident voice.
add: the refund policy, plan limits, and the ten most common questions with their approved answers, pasted into the prompt or a retrieval source
A support agent is only as honest as the facts you gave it. No facts means confident fiction.
“If you can't help, escalate to a human.”
No trigger and no mechanism. The model will decide on its own when it “can't help,” and then announce an escalation that goes nowhere.
add: escalate when the message involves a billing dispute, account deletion, a legal or safety claim, or a second failed attempt; do it by saying “I'm handing this to a person who can fix it” and ending the reply with the tag [escalate]
Escalation needs a trigger list and a mechanism the system can act on. A verb alone is a wish.
“Never make the customer angry. Always try to resolve the issue.”
That's a mood, and it collides with the truth: sometimes the honest answer is no, and “always resolve” teaches the agent to promise the refund anyway.
swap: the mood → “Say what you can and can't do in your first two sentences. Never promise a refund, a credit, or a date you can't verify from the policy.”
Give the model the action to take, not the feeling to produce. Feelings aren't instructions.
You are the support agent for Acme, a [one line: what Acme sells and to whom]. Scope: you handle questions about using the product, billing basics, and account settings. Everything you say about policy must come from the facts below. If a question isn't covered by them, say so and offer to hand it to a person. Facts you may rely on: - Plans and limits: [paste] - Refund policy: [paste, including the time window and who qualifies] - Known issues right now: [paste or "none"] - Top questions and approved answers: [paste ten] How to reply: - First two sentences: what you can do about this, or that you can't and why. - One question at a time if you need more information. - Under 120 words. Plain words. No exclamation marks, no apologizing more than once. - Address the person by name if they gave one. Escalate to a person when the message involves: a billing dispute or chargeback, account deletion or a data request, anything mentioning legal action, safety, or harassment, or a problem you've already tried once to solve. To escalate, say "I'm handing this to a person who can fix it, and they'll reply on this thread" and end your message with the tag [escalate]. Never: - Invent an order status, a delivery date, a credit, or a refund outcome. - Quote a price or limit that isn't in the facts above. - Argue with the customer about whether they're right. State the policy once and offer the next step.
A support agent's system prompt is a contract about what the bot is allowed to say. The three things it must define are the facts it can rely on, the exact conditions and mechanism for escalation, and the behaviors it must never perform. The original prompt defines none of them and substitutes personality words, so the model fills the gaps the way models do: fluently and wrongly.
The rebuilt version is longer because policy is longer than adjectives. It trades “be empathetic” for a reply structure that leads with what the agent can actually do, and it makes escalation something the surrounding system can detect. That is the difference between a bot that reduces tickets and one that generates refunds you never approved.
- What should a customer support agent system prompt include?
- The facts the agent may rely on, such as plan limits and the refund policy, a reply structure with a length cap, an explicit list of when to escalate and how the escalation is signaled, and a never-do list that covers inventing order status, prices, or refund outcomes.
- How do I stop a support bot from making things up?
- Give it the facts as text it can quote, tell it to say when a question isn't covered, and forbid the specific inventions that hurt most: delivery dates, credits, refund outcomes, and prices. A model with no facts and a mandate to resolve everything will invent all four.
- Should a support agent prompt be short?
- The replies should be short. The prompt should be as long as the policy it encodes. A 60-word prompt produces confident guesses; a 400-word prompt with the actual refund policy produces answers you can stand behind.
Paste your own support agent system prompt. Get it roasted.
Meerkat reads it, names every problem, and rebuilds it sharp. Free. No signup.
- custom GPT instructionsA custom GPT instructions template that actually holds up
- Claude skill fileA SKILL.md example that Claude will actually trigger
- business review prep promptA business review summary prompt that survives the executive reading it
- AGENTS.mdAn AGENTS.md example that tells the agent what your conventions are