Prompt patterns that survive contact with customers.

Not clever one-liners. These are the structural patterns that keep an agent stable under real traffic: how to phrase an escalation trigger so it fires reliably, how to make a refusal feel like service, how to force structured output that an integration can parse without a fallback parser.

The patterns

• Escalation trigger blocks — explicit conditions rather than confidence thresholds.
• Objection handling — acknowledge, reframe, offer, without arguing.
• Refusal framing — decline and route, in one sentence, without apology loops.
• Structured extraction — schemas that survive a caller who answers out of order.
• Confirmation patterns — read-back that catches wrong addresses before dispatch.

Why explicit beats implicit

Confidence-threshold escalation sounds elegant and behaves unpredictably, because a model can be confidently wrong. Explicit trigger conditions — the caller said refund, legal, complaint, or asked for a person by name — fire reliably and can be tested in an eval suite. Boring, and correct.

How to adapt them

Replace the example phrasing with sentences from your own recorded calls. The structure carries over between businesses; the language does not, and template language is the single clearest signal to a caller that they are talking to something bought rather than built.