# Prompt: Draft a Fair Customer Complaint Reply

> A reusable prompt that turns a pasted complaint plus your remediation limits into a review-ready reply draft, an issue list and a self-check.

## Install

Copy the content below into your project:

# Prompt: Draft a Fair Customer Complaint Reply

A reusable prompt that turns a pasted complaint plus your remediation limits into a review-ready reply draft, an issue list and a self-check.

## Start here

This is a plain prompt you paste into an ordinary AI chat that accepts text. No terminal, API or account setup is needed. Nothing is sent, refunded or credited — you only get a draft for a human to review.

Paste, in this order:

1. The customer's own words, in full (keep order numbers and timestamps).
2. FACTS_I_CONFIRM — what you personally verified. Write "none verified" if nothing is confirmed.
3. REMEDIATION_LIMITS — exactly what you may offer, including caps (for example: "replacement allowed up to 30 days", "no goodwill credit without manager approval").
4. TONE — one line, e.g. "plain, accountable, no corporate hedging".
5. CHANNEL_AND_LENGTH — e.g. "email, under 250 words".

Then paste the prompt itself. If the complaint text or your limits are empty, the prompt asks you to stop and supply them instead of inventing policy.

Check the output before using it: does every numbered grievance match the customer's own framing; is each row marked supported, contradicted or unverified; does every offer trace back to your stated limits; does the self-check honestly report pass or fail on all four checks; and is every guess tagged "assumed"? Ask a colleague to confirm anything that changes money owed before anything is sent.

## What it does

It produces, in plain Markdown: an issue list table (number | grievance | fact status | disposition | approval needed), a draft letter, missing inputs and assumptions, self-check findings, and a final revised draft. The included worked example is fictional and labelled as such.

## Prerequisites, permissions and limits

- You supply the complaint text and your authorization limits; the prompt has no access to accounts, order systems, billing or email.
- You may not claim the prompt sent, logged, credited or escalated anything.
- Legal liability, employment matters and regulated refund rules should go to a human.
- Keep customer data within your workplace rules; do not paste personal data into a tool you are not permitted to use.
- source reviewed; runtime not tested.

## FAQ

**Why does it refuse to fill in a missing policy?** Because inventing limits could promise money you are not allowed to spend. Empty limits mean the prompt stops and asks.

**Can I use the answer as-is?** No. It is preparation only; a human checks the assumptions and the flagged items before sending.

## Source

Original TokRepo prompt (CC BY 4.0), category: office. Reference context: [ChatGPT release notes](<https://help.openai.com/en/articles/6825453-chatgpt-release-notes>), reviewed 2026-10-04. External reference material keeps its own rights.

## Complete reusable prompt

You are drafting a customer complaint response letter for a human to review before anything is sent or credited. You have no access to accounts, order systems, billing, or email; you cannot apply refunds, issue credits, or send messages. Your only job is to produce a draft plus a check list from the material below.

INPUTS (fill these in before running):
1. COMPLAINT_TEXT — the customer's own words, pasted in full (keep original wording and any timestamps or order references).
2. FACTS_I_CONFIRM — what you, the responder, have verified as true (order date, delivery status, staff involved, prior replies). Write "none verified" if nothing is confirmed yet.
3. REMEDIATION_LIMITS — exactly what you are authorized to offer, and what you are not. Include hard caps (for example: "replacement allowed up to X", "refund not allowed beyond Y days", "no goodwill credit without manager approval").
4. TONE — one line, e.g. "plain, accountable, no corporate hedging".
5. CHANNEL_AND_LENGTH — e.g. "email, under 250 words" or "public comment reply, short".

WHAT TO DO:
Step 1 — Issue list. Extract every distinct grievance the customer raises, one per line, in the customer's own framing, numbered. Do not merge two different complaints into one line. If the complaint implies a grievance without stating it (for example, anger with no named cause), record it as "inferred" and mark it.
Step 2 — Fact check column. For each numbered issue, state what FACTS_I_CONFIRM supports: "supported", "contradicted", or "unverified". Never upgrade "unverified" to "supported" on your own.
Step 3 — Disposition per issue. Using only REMEDIATION_LIMITS: mark "in limits", "needs approval" (give the exact approval needed), or "cannot offer" (state the stated limit as the reason). Offer nothing outside those limits, and invent no policy, no timeline, and no compensation.
Step 4 — Draft the letter. Structure: (a) one sentence acknowledging the specific problem, (b) a short factual account drawn only from FACTS_I_CONFIRM (skip or label anything unverified), (c) the offer for each issue that is within limits, stated plainly, (d) what the customer must do next, if anything, (e) one clear closing line. Match TONE and CHANNEL_AND_LENGTH. Plain sentences, no flattery, no promises about future policy or speed you cannot guarantee.
Step 5 — Missing inputs. List what you needed but did not have, and what you assumed in its place, each tagged "assumed". Any assumption that could change money owed or policy interpretation must be flagged as requiring human confirmation before sending.
Step 6 — Self-check. Re-read your own draft and report: (i) every claim in the letter not traceable to FACTS_I_CONFIRM, (ii) every offer not traceable to REMEDIATION_LIMITS, (iii) any sentence that sounds like an admission of legal fault or a guarantee, (iv) any issue from Step 1 that the letter silently ignores. Then give pass/fail on each of the four checks and fix failures in a revised draft.

OUTPUT FORMAT (plain Markdown, in this order):
- Issue list table: number | customer's grievance | supported/contradicted/unverified | disposition | approval needed
- Draft letter
- Missing inputs and assumptions (tagged)
- Self-check findings with pass/fail per check
- Final revised draft

WORKED FICTIONAL EXAMPLE (illustration only; not real data):
COMPLAINT_TEXT: "My tote bag arrived with the strap torn. I emailed on the 3rd and nobody replied. I want a replacement and an apology."
FACTS_I_CONFIRM: delivery confirmed on the 2nd; no reply on record.
REMEDIATION_LIMITS: replacement allowed within 30 days; apology allowed; no refunds beyond 30 days; compensation credits need manager approval.
Expected shape: Issue 1 "torn strap" — supported — in limits → replacement offered. Issue 2 "no reply to 3rd" — supported — in limits → apology offered. Then the draft, the assumption tag for the missing order number, and the self-check. Task-specific checks on this example: FAIL if the letter promises a refund or a deadline not in REMEDIATION_LIMITS; FAIL if it states "we shipped on the 2nd" when FACTS_I_CONFIRM says delivered, not shipped; FAIL if issue 2 is dropped from the letter; FAIL if any unconfirmed fact appears without an "assumed" tag.

BOUNDARIES: this is preparation only. Do not claim to have sent, logged, credited, or escalated anything. Do not advise on legal liability, employment consequences, or regulated refund rules; flag those to a human instead. If COMPLAINT_TEXT or REMEDIATION_LIMITS is empty, stop and ask for it rather than filling the gap with plausible policy.

## References and reuse

- [ChatGPT release notes](https://help.openai.com/en/articles/6825453-chatgpt-release-notes) · Reviewed 2026-10-04

Original TokRepo prompt · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Reference documents retain their own rights.

---

<!-- ZH -->

# 提示词：起草公平的客户投诉回复信

一套可复用提示词，把粘贴的投诉原文与你的补救权限，转化为可供人工审核的回复草稿、问题清单和自检结果。

## 开始使用

这是一段普通提示词，粘贴到任何接受文本的 AI 对话里即可使用。不需要终端、API 或账号配置。它不会发送、退款或补偿任何东西，你得到的只是供人工审核的草稿。

请按顺序粘贴：

1. 客户的原始表述全文（保留订单号和日期时间）。
2. FACTS_I_CONFIRM —— 你亲自核实为真的内容。若尚无核实，写“none verified”。
3. REMEDIATION_LIMITS —— 你被授权提供的全部内容及上限（例如：“30 天内可换货”“无经理批准不得给善意补偿”）。
4. TONE —— 一行语气说明，例如“plain, accountable, no corporate hedging”。
5. CHANNEL_AND_LENGTH —— 例如“email, under 250 words”。

然后再粘贴提示词本身。如果投诉文本或你的权限为空，提示词会要求你先补齐，而不是自行编造政策。

使用前检查输出：每条编号的问题是否忠实于客户自己的说法；每条是否标为 supported、contradicted 或 unverified；每个承诺是否都能追溯到你的权限；自检是否对四项检查诚实给出 pass 或 fail；每一处猜测是否标为 assumed。任何涉及金额的猜测，发送前都请同事确认。

## 它能做什么

它以纯 Markdown 输出：问题清单表格（编号 | 客户诉求 | 事实状态 | 处置 | 所需审批）、回复草稿、缺失信息与假设、自检结果，以及最终的修订草稿。文中自带的示例为虚构内容并已标注。

## 前提、权限与限制

- 由你提供投诉原文和授权范围；提示词无法访问账户、订单系统、账单或邮箱。
- 不得声称该提示词已发送、记录、补偿或上报任何内容。
- 法律责任、雇佣问题和受监管的退款规则，应交由人工处理。
- 请遵守单位的客户数据规定；不要把个人信息粘贴到你无权使用的工具中。
- source reviewed; runtime not tested（已审阅来源；未做运行测试）。

## 常见问题

**为什么它拒绝补全缺失的政策？** 因为编造权限可能承诺你无权支出的钱。权限为空时，提示词会停下并询问。

**可以直接使用结果吗？** 不可以。它只做准备工作；发送前须由人工核对假设和已标记事项。

## 来源

TokRepo 原创提示词（CC BY 4.0），类别：办公。参考背景：[ChatGPT release notes](<https://help.openai.com/en/articles/6825453-chatgpt-release-notes>)，审阅于 2026-10-04。外部参考资料保留其自身权利。

## 完整可复制提示词

你正在起草一封客户投诉回复信，供人工在任何内容发送或补偿之前审核。你无权访问账户、订单系统、账单或电子邮件；你无法执行退款、发放补偿或发送消息。你唯一的任务是根据以下材料生成一份草稿和一份检查清单。

输入（运行前填写）：
1. COMPLAINT_TEXT —— 客户的原话，全文粘贴（保留原始措辞及任何时间戳或订单引用）。
2. FACTS_I_CONFIRM —— 你，即回复方，核实为真的内容（下单日期、配送状态、涉及员工、此前的回复）。若尚无任何核实，写“none verified”。
3. REMEDIATION_LIMITS —— 你被授权提供的准确范围，以及不被允许提供的范围。包括硬性上限（例如：“replacement allowed up to X”“refund not allowed beyond Y days”“no goodwill credit without manager approval”）。
4. TONE —— 一行语气说明，例如“plain, accountable, no corporate hedging”。
5. CHANNEL_AND_LENGTH —— 例如“email, under 250 words”或“public comment reply, short”。

操作步骤：
第 1 步 —— 问题清单。提取客户提出的每一项独立诉求，每行一项，忠实于客户自己的说法，编号。不要将两项不同的投诉合并为一行。如果投诉隐含某项诉求但未明说（例如愤怒但未指明原因），将其记录为“inferred”并标注。
第 2 步 —— 事实核查列。对每个编号问题，说明 FACTS_I_CONFIRM 支持什么：“supported”“contradicted”或“unverified”。绝不自行将“unverified”升级为“supported”。
第 3 步 —— 每个问题的处置。仅使用 REMEDIATION_LIMITS：标注“in limits”“needs approval”（给出所需的确切审批）或“cannot offer”（以所述限制作为理由）。不得提供超出这些范围的任何内容，不得编造任何政策、任何时限或任何补偿。
第 4 步 —— 起草信件。结构：(a) 一句话承认具体问题，(b) 仅依据 FACTS_I_CONFIRM 的简短事实陈述（跳过或标注任何未经核实的内容），(c) 对每个在权限范围内的问题给出承诺，平实陈述，(d) 客户接下来需要做什么（如有），(e) 一句明确的结尾。匹配 TONE 和 CHANNEL_AND_LENGTH。句子平实，不奉承，不对未来政策或你无法保证的速度作出承诺。
第 5 步 —— 缺失输入。列出你需要但没有的内容，以及你替代性假设的内容，每项标注“assumed”。任何可能改变应付金额或政策解释的假设，必须标记为发送前需人工确认。
第 6 步 —— 自检。重读你自己的草稿并报告：(i) 信中每项无法追溯到 FACTS_I_CONFIRM 的陈述，(ii) 每项无法追溯到 REMEDIATION_LIMITS 的承诺，(iii) 任何听起来像承认法律过错或作出保证的句子，(iv) 第 1 步中任何被信件悄然忽略的问题。然后对四项检查分别给出 pass/fail，并在修订草稿中修正未通过项。

输出格式（纯 Markdown，按此顺序）：
- 问题清单表格：编号 | 客户诉求 | supported/contradicted/unverified | 处置 | 所需审批
- 回复草稿
- 缺失信息与假设（已标注）
- 自检结果，每项检查含 pass/fail
- 最终修订草稿

虚构示例（仅供说明；非真实数据）：
COMPLAINT_TEXT: “My tote bag arrived with the strap torn. I emailed on the 3rd and nobody replied. I want a replacement and an apology.”
FACTS_I_CONFIRM: delivery confirmed on the 2nd; no reply on record.
REMEDIATION_LIMITS: replacement allowed within 30 days; apology allowed; no refunds beyond 30 days; compensation credits need manager approval.
预期形态：问题 1“torn strap”——supported——in limits → 提供换货。问题 2“no reply to 3rd”——supported——in limits → 提供道歉。然后是草稿、针对缺失订单号的假设标注，以及自检。对本示例的针对性检查：如果信件承诺了 REMEDIATION_LIMITS 中不存在的退款或期限，则 FAIL；如果 FACTS_I_CONFIRM 说是 delivered 而非 shipped，而信件却称“we shipped on the 2nd”，则 FAIL；如果问题 2 被信件遗漏，则 FAIL；如果任何未经确认的事实出现时未带“assumed”标注，则 FAIL。

边界：这仅是准备工作。不要声称已发送、记录、补偿或上报任何内容。不要就法律责任、雇佣后果或受监管的退款规则提供建议；应将这些标记出来交由人工处理。如果 COMPLAINT_TEXT 或 REMEDIATION_LIMITS 为空，应停止并索取它，而不是用看似合理的政策来填补空白。

## 参考资料与复用

- [ChatGPT release notes](https://help.openai.com/en/articles/6825453-chatgpt-release-notes) · Reviewed 2026-10-04

TokRepo 原创提示词 · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)。参考资料保留各自原有权利。

---
Source: https://tokrepo.com/en/workflows/prompt-draft-fair-customer-complaint-reply-7a450d2c
Author: Prompt Lab