On a Tuesday I will not forget, a customer replied to my carefully written support email with one line: "Did a human actually read my message, or did a bot send this?" I had read every word of his message. I had spent eleven minutes on the reply. And he thought a machine wrote it, because I opened with "We apologise for any inconvenience this may have caused." He had just lost two days of work to a sync bug, and I answered him with a phrase printed on the inside of a thousand corporate coffins.
That moment broke something useful in me. I realized my replies were technically polite and emotionally absent. The words were correct and the human was missing. I had been measuring my own support quality by whether I sounded professional, when the only metric that mattered to the person on the other end was whether I sounded like I cared.
Template empathy signals the opposite of empathy
Here is the cruel mechanics of it: "sorry for any inconvenience" tells an angry person you have NOT engaged with their specific situation. It is the linguistic equivalent of a customer service rep reading off a card while looking at their phone. When someone is furious, the single most de-escalating move available to you is to name their actual problem back to them, accurately, so they know you understood it. A template physically cannot do that, because a template does not know what happened to them.
I was treating empathy as a tone to apply at the end, like a coat of varnish. It is not. It is the first sentence, and it has to be specific, or it is worthless. I went back and read fifty of my own recent replies, and every single one of them opened with a stock phrase that could have been pasted onto any ticket in the queue. Interchangeable openings to people whose problems were anything but interchangeable. No wonder they thought a bot wrote them — functionally, a bot could have.
Writing under pressure without melting
The reason I leaned on stock phrases was simple: when a customer is yelling in all caps, my brain panics and reaches for the nearest safe-sounding words. That is exactly when I would over-promise too, blurting out a refund I had no authority to give just to make the shouting stop. The fear and the over-promising were the same reflex wearing two coats. So I started running these tickets through the De-escalate a Furious Customer Without Over-Promising prompt on Claude Sonnet 4.6. I paste the angry message, the confirmed facts, and crucially the exact list of what I am actually allowed to offer.
Claude is genuinely good at this because the prompt hands it light XML tags that separate the customer's emotion from the confirmed facts from the authorized offers. It opens with the real problem in plain human language, takes the right amount of ownership, states only what I gave it as fact, and lands on the one concrete next step I can actually deliver. It reads like a person who cared, because the structure forced me to engage with the specifics before a single word got written.
I will be honest about the one tweak I had to make. The first few outputs were a touch too composed — almost too emotionally fluent, in a way that could read as performed for a deeply angry customer. So I added a single line to my facts input reminding it to keep the acknowledgment short and let the action carry the weight. After that, the replies hit the exact register I wanted: brief warmth, then a real next step. The lesson was that the prompt is a floor of quality I cannot drop below, not a ceiling I am forbidden from adjusting.
- Replies now open with the real problem named back, not a generic apology
- The 'offer only what you are authorised to offer' rule stopped me caving and over-promising under fire
- When a fact is missing, it asks one clarifying question instead of confidently inventing one
- My escalation-to-supervisor rate on hot tickets dropped by about a third in two months
The honesty rule is the whole game
The hard rule I value most is the one against promising anything outside my authorized list. When a customer is angry, the instinct is to placate them with a promise, and a promise you cannot keep is just a delayed, bigger fire. The prompt holds the line: empathy in the words, honesty in the offer. That combination de-escalates far better than a refund I would have to claw back tomorrow.
I pair this with my Summarize a Long Ticket Thread and Surface Every Promise prompt when an angry ticket has already bounced between three agents, so I know what was promised before I reply. But for the raw, incoming fury, this one is my reflex now. The summary tells me what I am walking into; this prompt helps me walk in without flinching or over-promising.
One more honest caveat, because I promised myself I would never write a glowing blog with no downside. This prompt is a reply drafter, not a judgment replacement. It will faithfully say only what I authorize, which means if I feed it a stingy offer, it will deliver that stingy offer beautifully — and a beautifully worded inadequate resolution is still inadequate. The prompt fixed my words. It could not fix a policy that genuinely deserved the customer's anger. On the tickets where the customer was right and we were wrong, my job was still to go fight for a better offer first, then come back and let the prompt deliver it like a human.
Grab the De-escalate a Furious Customer Without Over-Promising prompt on Prompt Dock and run it on your own scariest open ticket. Nobody has asked me if a bot wrote my replies since that Tuesday, and the difference was not a better template. It was finally saying the true, specific thing first.