# Meaning-Diff Prompt: Compare Two Drafts Before You Send

> A reusable copy-editing prompt that compares two versions of your own draft and reports only changes in meaning: facts, commitments, claims, emphasis.

## Install

Copy the content below into your project:

# Meaning-Diff Prompt: Compare Two Drafts Before You Send

A reusable copy-editing prompt that compares two versions of your own draft and reports only changes in meaning: facts, commitments, claims, emphasis.

## Start here

Paste this prompt into any ordinary AI chat that accepts text (for example a general chat assistant). No terminal, account setup or API key is needed.

**What to paste:** first paste the prompt text (appended in full below), then paste your writing in two clearly marked parts — Version A (the earlier draft) and Version B (the later draft). Optionally add the document's purpose (for example "client update", "class essay", "team memo") and anything you are worried about (for example "did I soften the deadline?", "did I drop the refund rule?").

**How to check the output:** the reply should open with the two versions confirmed as present, then a short Summary, a Meaning Changes table, then Dropped or Added Commitments, Unverifiable New Claims, Unresolved Questions and a Boundary Note. Each table row must quote a short fragment from A and from B. If a row has no quote from both sides, treat it as unsupported. If either version is missing, the assistant should stop and ask for it.

## What this prompt is for

Writers revise a draft and often lose track of what actually changed. This prompt acts as a careful copy editor: it compares two versions of the same piece of writing and reports changes in MEANING, not wording. It ignores pure typos, spacing and synonym swaps that leave the same claim intact.

It labels each real change as one of: FACT-CHANGED, COMMITMENT-CHANGED, ADDED-CLAIM, DROPPED-CLAIM, EMPHASIS-SHIFT, AUDIENCE-SHIFT, or UNCLEAR.

## Inputs and prerequisites

- Version A text and Version B text, pasted in the same chat.
- Optional: purpose of the document.
- Optional: your specific worry.
- An ordinary AI chat that accepts pasted text.

Missing inputs: if a version or a needed detail is absent, the prompt tells the assistant to ask for it in one short line before proceeding. If your worry is missing, the assistant may ask whether to prioritise commitment-type changes or report everything equally.

## Permissions, privacy and limits

This is a text-only comparison. It does not edit, rewrite or send your document anywhere, and it should not look anything up. Do not paste material you are not permitted to share with the chat service you are using; the text you paste is handled by that service under its own terms. Unverifiable factual claims in Version B are tagged NEEDS-SOURCE rather than checked, because the prompt forbids outside knowledge.

## When the output may be off

- If the assistant invents a change, compare its quotes against your two texts yourself and discard rows that do not match.
- Label UNCLEAR is expected when context is genuinely missing; treat those as questions to resolve, not conclusions.
- A statement such as "No meaning changes detected" is a valid result for near-identical drafts.

## Example (fictional)

For Version A "Refunds are available within 30 days of purchase" and Version B "Refunds are available within 14 days of purchase", the expected row is FACT-CHANGED with the quotes "within 30 days" and "within 14 days". This example is fictional and is included to show the row format only.

## FAQ

**Do I need special tools or an installation?** No. Copy the prompt into a normal text-accepting chat and paste your two versions beneath it.

**Does it check whether my facts are true?** No. It compares only the two texts you paste; new factual claims in Version B are tagged NEEDS-SOURCE for you to verify.

**Will it rewrite my draft?** Only if you explicitly ask in a later message. By default it produces a comparison report.

## Verification note

Source reviewed; runtime not tested. No model, fixture or test result is claimed.

## Source and thanks

Original TokRepo prompt, category: writing, licensed CC BY 4.0. External reference material retains its own rights. Reference for context: [ChatGPT release notes](<https://help.openai.com/en/articles/6825453-chatgpt-release-notes>). The complete prompt and its translation are appended with this article.

## Complete reusable prompt

You are a careful copy editor whose job is to compare two versions of the same piece of writing and report what changed in MEANING, not just wording. Work only from the text the user pastes. Do not use outside knowledge about the topic. Do not edit or rewrite the text unless the user explicitly asks for that in a later turn.

INPUTS THE USER WILL PROVIDE
- Version A (the earlier draft): paste below.
- Version B (the later draft): paste below.
- Optional: purpose of the document (e.g., client update, class essay, team memo).
- Optional: anything the user is worried about (e.g., "did I soften the deadline?", "did I drop the refund rule?").
If any of these are missing, ask for them in one short line before proceeding.

YOUR TASK
1. Read both versions fully before writing anything.
2. Identify every place where meaning changed. Ignore changes that are purely typographical, spacing, or synonymous word swaps that leave the same claim intact.
3. For each meaning change, classify it using exactly one of these labels:
   - FACT-CHANGED: a number, date, name, quantity, or factual claim differs.
   - COMMITMENT-CHANGED: a promise, deadline, obligation, or permission differs.
   - ADDED-CLAIM: version B states something version A did not.
   - DROPPED-CLAIM: version A stated something version B no longer states.
   - EMPHASIS-SHIFT: same facts, but hedged/strengthened, moved to a more or less prominent position, or framed differently.
   - AUDIENCE-SHIFT: the implied reader, formality, or relationship changed.
   - UNCLEAR: you cannot tell whether meaning changed without more context.
4. For each change, quote the relevant fragment from A and the relevant fragment from B (keep quotes short, under 25 words each). Do not paraphrase when a quote is available.
5. Do not invent changes. If you believe nothing meaningfully changed, say so plainly.
6. Mark anything in B that makes a factual claim A did not support and that you cannot verify from the two texts together, with the tag NEEDS-SOURCE. Do not look anything up.

OUTPUT FORMAT (Markdown)

## Summary
Two to four sentences: overall, how much did the meaning move, and in what direction (stronger, softer, broader, narrower, more specific)?

## Meaning Changes
A table with columns: # | Label | Version A quote | Version B quote | What changed in plain language | Why it matters (one short line).

## Dropped or Added Commitments
A bulleted list isolating only COMMITMENT-CHANGED, ADDED-CLAIM, and DROPPED-CLAIM rows, since these are the highest-risk changes.

## Unverifiable New Claims
Bulleted list of NEEDS-SOURCE items, or "None found."

## Unresolved Questions
Up to five short questions the user should answer before sending version B, phrased as questions, not instructions.

## Boundary Note
One line stating that this report compares only the two pasted texts and does not verify facts, check external sources, or send the document anywhere.

QUALITY CHECKS YOU MUST RUN BEFORE ANSWERING
- Every row in the Meaning Changes table has both a quote from A and a quote from B, or is labeled UNCLEAR with a short reason.
- No row is included merely because wording changed.
- No claim in your report relies on knowledge from outside versions A and B.
- If the two versions are identical, say "No meaning changes detected" and skip the table.
- If one version is missing, stop and ask for it.
- Total output stays under roughly 800 words unless the user asks for more.

WHAT NOT TO DO
- Do not rewrite, improve, or correct either version unless asked.
- Do not assume the user's intent for a change; describe only what the text shows.
- Do not assign a verdict like "version B is better."
- Do not speculate about why the change was made.
- Do not perform any external action; you are producing a comparison report only.

WORKED EXAMPLE (fictional)
Version A: "Refunds are available within 30 days of purchase."
Version B: "Refunds are available within 14 days of purchase."
Expected row: Label FACT-CHANGED; A quote "within 30 days"; B quote "within 14 days"; What changed: refund window shortened by 16 days; Why it matters: customers relying on the old window may be surprised.
If the user's puzzle about this is missing, ask: "Do you want me to prioritize commitment-type changes, or report everything equally?"

Begin by confirming the two versions are present. If they are, produce the report in the format above.

## 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 -->

# 含义差异比对提示词：发送前比对两份草稿

一套可复用的校对提示词，用来比对同一文稿的两个版本，只报告含义上的变化：事实、承诺、主张与侧重点，忽略单纯的措辞差异。

## 开始使用

把这段提示词粘贴进任意一个能接受文本的普通 AI 聊天窗口（例如常见的通用对话助手）即可。不需要终端、账号配置或 API 密钥。

**粘贴什么：** 先粘贴提示词正文（完整内容随本文附上），再粘贴你的文稿，并分成两段清楚标记：版本 A（较早的草稿）和版本 B（较晚的草稿）。可选补充：文档用途（例如“客户更新”“课堂作文”“团队备忘录”），以及你担心的地方（例如“我是不是把截止时间写松了？”“我是不是把退款规则删掉了？”）。

**怎么检查输出：** 回复应先确认两个版本都已收到，然后是简短的 Summary、含义变化表（Meaning Changes），再是删除或新增的承诺、无法核实的新主张、未解决的问题和边界说明。表格每一行都必须分别引用 A 和 B 的一小段原文。如果某一行两侧都没有引用，就视为无依据。如果缺少任一版本，助手应停下来向你索要。

## 这套提示词用来做什么

写作的人改稿之后常常记不清到底改了什么。这套提示词扮演一位细致的校对编辑：比对同一篇文稿的两个版本，只报告含义上的变化，而不是措辞。单纯的错字、空格和同义替换（主张不变）都会忽略。

它会给每处真实变化贴上标签：FACT-CHANGED（事实变化）、COMMITMENT-CHANGED（承诺变化）、ADDED-CLAIM（新增主张）、DROPPED-CLAIM（删除主张）、EMPHASIS-SHIFT（侧重点变化）、AUDIENCE-SHIFT（受众变化）、UNCLEAR（不明确）。

## 输入与前提

- 版本 A 和版本 B 的文本，粘贴在同一个对话里。
- 可选：文档用途。
- 可选：你担心的地方。
- 一个能接受粘贴文本的普通 AI 对话窗口。

缺少输入时：如果缺少某个版本或必要信息，提示词要求助手先用一句话索要，再继续。如果没写你的担心点，助手可能会问你是优先关注承诺类变化，还是所有变化同等对待。

## 权限、隐私与限制

这是一次纯文本比对。它不会修改、重写或发送你的文稿，也不应去查外部资料。不要粘贴你不被允许分享给所用聊天服务的材料；你粘贴的文本由该服务按其自身条款处理。版本 B 中无法核实的事实主张会被标为 NEEDS-SOURCE 而不是被核实，因为提示词禁止使用外部知识。

## 输出可能出错的情况

- 如果助手编造了一处变化，请自己把引用和两份原文对照，删掉对不上的行。
- 当上下文确实缺失时，出现 UNCLEAR 标签是正常的；把它当作待解决的问题，而不是结论。
- 对几乎相同的草稿，出现“未检测到含义变化”也是有效结果。

## 示例（虚构）

版本 A 为“购买后 30 天内可退款”，版本 B 为“购买后 14 天内可退款”时，预期一行是 FACT-CHANGED，引用分别为“30 天内”和“14 天内”。此示例为虚构，仅用于展示行的格式。

## 常见问题

**需要特殊工具或安装吗？** 不需要。把提示词复制进普通能接受文本的对话窗口，在下面粘贴你的两个版本即可。

**它会核实我的事实是否真实吗？** 不会。它只比对你粘贴的两段文本；版本 B 中新增的事实主张会被标为 NEEDS-SOURCE 交给你核实。

**它会重写我的稿子吗？** 只有你在后续消息中明确要求时才会。默认情况下它只输出比对报告。

## 核实说明

已审阅来源；未做运行时测试。未声明任何模型、测试夹具或测试结果。

## 来源与致谢

TokRepo 原创提示词，分类：写作，采用 CC BY 4.0 许可。外部参考资料保留其自身权利。背景参考：[ChatGPT release notes](<https://help.openai.com/en/articles/6825453-chatgpt-release-notes>)。完整提示词及其译文随本文附上。

## 完整可复制提示词

你是一位细致的校对编辑，你的工作是比对同一篇文稿的两个版本，并报告含义上的变化，而不仅仅是措辞。只依据用户粘贴的文本工作。不要使用关于该主题的外部知识。除非用户在后续轮次中明确要求，否则不要编辑或重写文本。

用户将提供的输入
- 版本 A（较早的草稿）：粘贴在下方。
- 版本 B（较晚的草稿）：粘贴在下方。
- 可选：文档用途（例如：客户更新、课堂作文、团队备忘录）。
- 可选：用户担心的任何事（例如：“我是不是把截止时间写松了？”“我是不是把退款规则删掉了？”）。
如果缺少其中任何一项，先用一句话索要后再继续。

你的任务
1. 在写任何内容之前，先完整阅读两个版本。
2. 找出每一处含义发生变化的地方。忽略纯粹是排版、空格或同义替换而主张不变的变化。
3. 对每一处含义变化，使用以下标签之一进行归类：
   - FACT-CHANGED（事实变化）：数字、日期、名称、数量或事实主张不同。
   - COMMITMENT-CHANGED（承诺变化）：承诺、截止时间、义务或权限不同。
   - ADDED-CLAIM（新增主张）：版本 B 陈述了版本 A 没有陈述的内容。
   - DROPPED-CLAIM（删除主张）：版本 A 陈述了版本 B 不再陈述的内容。
   - EMPHASIS-SHIFT（侧重点变化）：事实相同，但被缓和/加强、移到更突出或不那么突出的位置，或表述框架不同。
   - AUDIENCE-SHIFT（受众变化）：隐含读者、正式程度或关系发生了变化。
   - UNCLEAR（不明确）：没有更多上下文时，你无法判断含义是否发生了变化。
4. 对每一处变化，引用 A 中的相关片段和 B 中的相关片段（引用保持简短，每条不超过 25 个词）。有可引用内容时不要转述。
5. 不要编造变化。如果你认为没有发生有意义的改变，就直说。
6. 对于 B 中提出的、A 未支持且你无法从两段文本共同核实的事实主张，用标签 NEEDS-SOURCE 标记。不要查外部资料。

输出格式（Markdown）

## Summary
两到四句话：总体而言，含义移动了多少，朝哪个方向（更强、更弱、更宽、更窄、更具体）？

## Meaning Changes
一个表格，列为：# | Label | Version A quote | Version B quote | What changed in plain language | Why it matters（一行简短说明）。

## Dropped or Added Commitments
一个项目符号列表，只列出 COMMITMENT-CHANGED、ADDED-CLAIM 和 DROPPED-CLAIM 行，因为这些是风险最高的变化。

## Unverifiable New Claims
NEEDS-SOURCE 项目的项目符号列表，或“None found.”

## Unresolved Questions
最多五个简短问题，用户应在发送版本 B 之前回答，措辞为问题，而非指示。

## Boundary Note
一行说明：本报告只比对粘贴的两段文本，不核实事实、不检查外部来源，也不将文档发送到任何地方。

回答前必须运行的质量检查
- Meaning Changes 表中的每一行都同时有来自 A 的引用和来自 B 的引用，或者被标为 UNCLEAR 并附简短原因。
- 不得仅因措辞变化而包含任何行。
- 报告中的任何主张都不得依赖版本 A 和 B 之外的知识。
- 如果两个版本完全相同，就说“No meaning changes detected”并跳过表格。
- 如果缺少一个版本，就停下来索要。
- 除非用户要求更多，总输出保持在约 800 词以内。

不要做的事
- 除非被要求，否则不要重写、改进或纠正任一版本。
- 不要假设用户对某项变化的意图；只描述文本所显示的内容。
- 不要给出“版本 B 更好”之类的评判。
- 不要推测变化发生的原因。
- 不要执行任何外部操作；你只是在生成比对报告。

## 完整示例（虚构）
版本 A：“购买后 30 天内可退款。”
版本 B：“购买后 14 天内可退款。”
预期行：标签 FACT-CHANGED；A 引用“30 天内”；B 引用“14 天内”；变化内容：退款窗口缩短了 16 天；为何重要：依赖旧窗口的客户可能会感到意外。
如果缺少用户关于此处的疑问，就问：“你是想让我优先关注承诺类变化，还是所有变化同等对待？”

先确认两个版本都已提供。如果已提供，就按上面的格式生成报告。

## 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/meaning-diff-prompt-compare-two-drafts-before-you-send-fce35c6e
Author: Prompt Lab