Skip to content
PromptsBuddy
All prompts
SupportintermediateStaff pickClaudeChatGPT

The support reply that de-escalates without grovelling

Drafts a response to an angry customer that acknowledges the real complaint, commits to something specific, and never apologises for things that aren't your fault.

DineshDinesh · Head of supportJun 4, 2026Updated Aug 11, 2026v2.0
The prompt· 0/3 variables filled
~317 tokens
Draft a reply to this customer.

CUSTOMER MESSAGE:
{{MESSAGE}}

WHAT I ACTUALLY KNOW:
{{FACTS}}

WHAT I CAN OFFER:
{{AUTHORITY}}

Before writing, answer these three internally and show them as a short block
labelled ANALYSIS:
1. What is the customer's stated complaint?
2. What is the underlying one? (Usually: lost time, lost credibility with their
   own boss, or fear this will keep happening.)
3. What outcome would make them consider this resolved?

Then write the reply. Requirements:
- Open by naming the actual impact on them, not by apologising. One sentence.
- State what happened in plain language. No passive voice hiding the actor.
- Commit to exactly one specific next step with a date or a named person.
- Never promise anything outside WHAT I CAN OFFER. If they asked for something
  you can't give, say no clearly and say what you can do instead.
- No "we sincerely apologise for any inconvenience". No "we value your feedback".
- Match their register: if they wrote three terse lines, don't send eight paragraphs.
- End without a question unless you genuinely need an answer to proceed.

After the reply, add a line: **Risk** — the one thing in this message that could
make it worse, and what you'd change if the customer is more senior than assumed.

Fill in the blanks

Run it in:ClaudeChatGPT

What goes in each blank

MESSAGE
The customer's message, verbatim.e.g. Third time this month your export has failed. We're evaluating alternatives.
FACTS
What actually happened, including anything you can't say publicly.e.g. Known bug in the CSV writer for >50k rows. Fix ships Thursday. Not disclosed publicly yet.
AUTHORITY
What you're actually allowed to offer.e.g. Can offer a credit up to one month. Cannot promise dates for unshipped work.

Notes from people who use it

  • The ANALYSIS block is for you, not the customer. Read it, then delete it before sending.
  • 'Never apologise first' surprises people. Naming the impact reads as understanding; a lead apology reads as a script.
  • If the model keeps sneaking in a promise you can't keep, tighten AUTHORITY — that field is doing the guardrail work.

Use this as a skill

A prompt is a skill with the serial numbers filed off. Install it once and your agent reaches for it on its own.

Paste this into ChatGPT, Claude Code, Codex, Cursor, or whatever you use:

Install the support-reply-that-de-escalates skill globally from https://promptsbuddy.com/prompts/support-reply-that-de-escalates/skill.md
Add to Cursor

.claude/skills/support-reply-that-de-escalates/SKILL.md

---
name: support-reply-that-de-escalates
description: "Drafts a response to an angry customer that acknowledges the real complaint, commits to something specific, and never apologises for things that aren't your fault. Use when the user asks for help with support tasks like support, customer, writing."
license: CC-BY-4.0
metadata:
  source: https://promptsbuddy.com/prompts/support-reply-that-de-escalates
  author: "Dinesh"
  version: "2.0"
---

# The support reply that de-escalates without grovelling

Drafts a response to an angry customer that acknowledges the real complaint, commits to something specific, and never apologises for things that aren't your fault.

## Inputs to collect first

Ask the user for anything below that they have not already given you. Do not
invent values for these.

- `MESSAGE` — The customer's message, verbatim. (e.g. Third time this month your export has failed. We're evaluating alternatives.)
- `FACTS` — What actually happened, including anything you can't say publicly. (e.g. Known bug in the CSV writer for >50k rows. Fix ships Thursday. Not disclosed publicly yet.)
- `AUTHORITY` — What you're actually allowed to offer. (e.g. Can offer a credit up to one month. Cannot promise dates for unshipped work.)

## Instructions

Draft a reply to this customer.

CUSTOMER MESSAGE:
{{MESSAGE}}

WHAT I ACTUALLY KNOW:
{{FACTS}}

WHAT I CAN OFFER:
{{AUTHORITY}}

Before writing, answer these three internally and show them as a short block
labelled ANALYSIS:
1. What is the customer's stated complaint?
2. What is the underlying one? (Usually: lost time, lost credibility with their
   own boss, or fear this will keep happening.)
3. What outcome would make them consider this resolved?

Then write the reply. Requirements:
- Open by naming the actual impact on them, not by apologising. One sentence.
- State what happened in plain language. No passive voice hiding the actor.
- Commit to exactly one specific next step with a date or a named person.
- Never promise anything outside WHAT I CAN OFFER. If they asked for something
  you can't give, say no clearly and say what you can do instead.
- No "we sincerely apologise for any inconvenience". No "we value your feedback".
- Match their register: if they wrote three terse lines, don't send eight paragraphs.
- End without a question unless you genuinely need an answer to proceed.

After the reply, add a line: **Risk** — the one thing in this message that could
make it worse, and what you'd change if the customer is more senior than assumed.

## Notes from people who use this

- The ANALYSIS block is for you, not the customer. Read it, then delete it before sending.
- 'Never apologise first' surprises people. Naming the impact reads as understanding; a lead apology reads as a script.
- If the model keeps sneaking in a promise you can't keep, tighten AUTHORITY — that field is doing the guardrail work.

---

The support reply that de-escalates without grovelling · by Dinesh · v2.0
From PromptsBuddy — https://promptsbuddy.com/prompts/support-reply-that-de-escalates
Licensed CC BY 4.0.

Every prompt is also available at /prompts/support-reply-that-de-escalates/skill.md — see all install options.

Why the analysis step is non-negotiable

An angry ticket has two complaints in it. The stated one ("the export failed") is usually cheap to address. The real one ("I told my team this tool was reliable") is what determines whether the customer stays. Ask the model to separate them before writing and the reply lands differently — it addresses the person, not the bug.

The apology trap

Support teams are trained to lead with an apology, and models are trained to match. The result is a reply that sounds identical to every automated one the customer has ever received, which reads as nobody actually looked at this. Naming the impact first — "three failed exports in a month is enough to make you question whether this is worth the trouble" — does more than any amount of sincerity language.

Guardrails that actually hold

AUTHORITY is not a formality. Without it, models offer refunds, ship dates, and escalations they have no business offering, and someone has to walk that back. Keep it narrow, keep it current, and treat any output that exceeds it as a bug in the prompt rather than a suggestion.

Try these next

WritingbeginnerStaff pick

Rewrite until it stops sounding like AI wrote it

A diagnostic-then-edit pass that names the specific tells in a draft — the tricolons, the hedges, the empty transitions — and fixes them without flattening your voice.

ClaudeChatGPT3 vars
RonyRony