# Evidence Comparison Prompt: Claims vs. Sources

> A reusable research prompt that turns pasted source excerpts into a claim table, agreement and disagreement notes, and follow-up questions.

## Install

Copy the content below into your project:

# Evidence Comparison Prompt: Claims vs. Sources

A reusable research prompt that turns pasted source excerpts into a claim table, agreement and disagreement notes, and follow-up questions.

## Quick Use

1. Copy the prompt text below.
2. Paste it into an ordinary AI chat that accepts text.
3. Replace the final line with your own question and two or more source excerpts, including titles, URLs and dates where you know them.
4. Send it. The reply should open with a claim table, then a short comparison, then three follow-up questions.

Use this prompt text:

Act as an evidence organizer for a non-specialist reader. I will paste two or more source excerpts, with their titles, URLs and dates if available, and a specific question. Use only those excerpts. Do not browse, fabricate a citation or use remembered facts as evidence. If the question or excerpts are missing, ask for them and stop.

Build a claim table with these columns: claim, which excerpt supports it, exact short supporting quote, source anchor, what the evidence actually establishes, and what remains unknown. Separate an author's assertion from evidence described in the excerpt. A repeated claim across several excerpts is not automatically independent confirmation. Note when sources appear to rely on the same original announcement or dataset.

Produce a short comparison of points of agreement, points of disagreement and missing information. Preserve dates and definitions. If the excerpts use different measures or populations, explain why the apparent comparison may be invalid. Do not rank sources by prestige alone, infer causation from correlation or turn absent evidence into evidence of absence.

End with three focused follow-up questions that would resolve the most important gaps and say what kind of primary evidence would help. These are research suggestions, not findings. Quote only small portions of the excerpts; summarize the rest in original language. Clearly label any hypothesis as an inference from the supplied text.

This is an organizational aid, not legal, medical or investment advice. Do not contact people, post conclusions or change any records. I must confirm significant conclusions against the original material.

Question and excerpts: paste them here with whatever source labels are actually known.

## Introduction

This is an original TokRepo prompt for general readers doing light research. It is an organizational aid, not a fact-checking service and not a substitute for reading your sources. Its value is that it keeps claims tied to the exact excerpts you supply, instead of letting a chat model fill gaps from memory.

The prompt asks for a claim table, a short comparison and three follow-up questions. It is written for people who already have a few excerpts and want to see what those excerpts do and do not support.

## Prerequisites

- No terminal, account or API setup is required.
- An ordinary AI chat that accepts pasted text.
- Two or more source excerpts you are allowed to paste, plus a clear question.
- Optionally, titles, URLs and dates for each excerpt.

## What You Need to Provide

| Input | Why it matters |
| --- | --- |
| A specific question | The prompt stops and asks for it if missing |
| Two or more excerpts | A single excerpt cannot be compared |
| Source labels | So the table can point to a named source |
| Titles, URLs, dates | The prompt preserves them where known |

Missing question: the prompt should ask you for it and stop. Missing excerpts: same. If you paste only one excerpt, expect weak or refused comparison rather than a table.

## Permissions and Limits

You are asking a chat model to organize text you supply. You are not authorizing it to browse, to contact anyone, to post conclusions or to change records. The prompt explicitly says not to do those things, but you remain responsible for what you send and how you use the reply.

Limitations to keep in mind:

- It works only from your excerpts. Absence of evidence in your excerpts is not evidence of absence.
- It cannot verify whether an excerpt itself is accurate.
- Excerpts that share one original announcement or dataset are not independent.
- The output is not legal, medical or investment advice.
- A confident-looking table is still a draft; confirm significant conclusions against the original material.

### Privacy

Pasting text into a chat may store or transmit it. Do not paste confidential, personal or licensed material unless you are allowed to. Redact names, contact details and anything you cannot share.

Documented setup instructions: source reviewed; runtime not tested.

## How to Check the Output

1. Does every row point to an excerpt you actually pasted?
2. Is the supporting quote short and does it appear in that excerpt?
3. Are author assertions separated from described evidence?
4. Are dates and definitions preserved rather than normalized?
5. Does it flag shared-origin sources?
6. Are the three follow-up questions actually questions, not conclusions?
7. Is any hypothesis labelled as an inference from your text?

## FAQ

### What if the model adds facts from its own memory?

Send the correction: use only the pasted excerpts, and mark anything else as unsupported. The prompt already forbids browsing and remembered facts, but models can drift; re-anchor with a short reminder and ask it to strike unsupported rows.

### Can I treat a repeated claim as confirmed?

The prompt instructs that repetition across excerpts is not automatically independent confirmation. Check whether the excerpts rely on the same original announcement or dataset before treating agreement as stronger evidence.

### What are the three follow-up questions for?

They point at the most important gaps and suggest what kind of primary evidence would help. They are research suggestions, not findings, and they are not a to-do list of actions to take.

### What if my excerpts disagree?

The comparison section should show points of agreement, points of disagreement and missing information. If the excerpts use different measures or populations, expect the prompt to explain why the apparent comparison may be invalid rather than pick a winner.

## Source and Thanks

Original TokRepo prompt, licensed CC BY 4.0. Reference context for a general-purpose AI workspace capability was drawn from the linked provider documentation dated as reviewed 2026-10-04; that reference is context only and does not imply any feature dependency. External reference material retains its own rights. Thanks to the provider documentation maintainers for the public reference.

## Complete reusable prompt

Act as an evidence organizer for a non-specialist reader. I will paste two or more source excerpts, with their titles, URLs and dates if available, and a specific question. Use only those excerpts. Do not browse, fabricate a citation or use remembered facts as evidence. If the question or excerpts are missing, ask for them and stop.

Build a claim table with these columns: claim, which excerpt supports it, exact short supporting quote, source anchor, what the evidence actually establishes, and what remains unknown. Separate an author's assertion from evidence described in the excerpt. A repeated claim across several excerpts is not automatically independent confirmation. Note when sources appear to rely on the same original announcement or dataset.

Produce a short comparison of points of agreement, points of disagreement and missing information. Preserve dates and definitions. If the excerpts use different measures or populations, explain why the apparent comparison may be invalid. Do not rank sources by prestige alone, infer causation from correlation or turn absent evidence into evidence of absence.

End with three focused follow-up questions that would resolve the most important gaps and say what kind of primary evidence would help. These are research suggestions, not findings. Quote only small portions of the excerpts; summarize the rest in original language. Clearly label any hypothesis as an inference from the supplied text.

This is an organizational aid, not legal, medical or investment advice. Do not contact people, post conclusions or change any records. I must confirm significant conclusions against the original material.

Question and excerpts: paste them here with whatever source labels are actually known.

## 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 -->

# 证据比对提示词：论点与来源对照

一条可复用的研究提示词，把粘贴的来源摘录整理成论点对照表、共识与分歧说明，以及后续追问问题。

## 快速使用

1. 复制下面的提示词文本。
2. 粘贴到能接受文本的普通 AI 聊天窗口中。
3. 把最后一行换成你自己的问题，以及两段或更多来源摘录，尽量附上标题、网址和日期。
4. 发送。回复应先给出论点对照表，再给出简短比较，最后给出三个追问问题。

使用以下提示词文本：

Act as an evidence organizer for a non-specialist reader. I will paste two or more source excerpts, with their titles, URLs and dates if available, and a specific question. Use only those excerpts. Do not browse, fabricate a citation or use remembered facts as evidence. If the question or excerpts are missing, ask for them and stop.

Build a claim table with these columns: claim, which excerpt supports it, exact short supporting quote, source anchor, what the evidence actually establishes, and what remains unknown. Separate an author's assertion from evidence described in the excerpt. A repeated claim across several excerpts is not automatically independent confirmation. Note when sources appear to rely on the same original announcement or dataset.

Produce a short comparison of points of agreement, points of disagreement and missing information. Preserve dates and definitions. If the excerpts use different measures or populations, explain why the apparent comparison may be invalid. Do not rank sources by prestige alone, infer causation from correlation or turn absent evidence into evidence of absence.

End with three focused follow-up questions that would resolve the most important gaps and say what kind of primary evidence would help. These are research suggestions, not findings. Quote only small portions of the excerpts; summarize the rest in original language. Clearly label any hypothesis as an inference from the supplied text.

This is an organizational aid, not legal, medical or investment advice. Do not contact people, post conclusions or change any records. I must confirm significant conclusions against the original material.

Question and excerpts: paste them here with whatever source labels are actually known.

## 简介

这是 TokRepo 的原创提示词，面向做轻量研究的一般读者。它是整理辅助工具，不是事实核查服务，也不能替代你自己阅读来源。它的价值在于把论点牢牢绑定在你提供的摘录上，而不是让聊天模型用记忆去填补空白。

提示词会要求一份论点对照表、一段简短比较和三个追问问题，适合那些已经拿到几段摘录、想看清这些摘录支持什么、又不支持什么的人。

## 前提条件

- 不需要终端、账号或 API 设置。
- 一个能接受粘贴文本的普通 AI 聊天窗口。
- 两段或更多你有权粘贴的来源摘录，以及一个明确的问题。
- 可选：为每段摘录提供标题、网址和日期。

## 你需要提供什么

| 输入 | 为什么重要 |
| --- | --- |
| 一个具体问题 | 缺失时提示词会先索要并停止 |
| 两段或更多摘录 | 只有一段无法比较 |
| 来源标签 | 对照表才能指向具名来源 |
| 标题、网址、日期 | 提示词会在已知时保留它们 |

缺少问题：提示词应当向你索要并停止。缺少摘录：同样处理。如果只粘贴一段摘录，应有心理准备得到的是薄弱比较或直接拒绝，而不是一张完整的对照表。

## 权限与限制

你是在请聊天模型整理你自己提供的文本，并没有授权它去浏览网页、联系他人、发布结论或改动任何记录。提示词已明确禁止这些行为，但你仍需对自己发送的内容和如何使用回复负责。

需要记住的限制：

- 它只依据你的摘录工作。摘录里没有证据，不等于事实不存在。
- 它无法验证摘录本身是否准确。
- 共用同一原始公告或数据集的摘录并不相互独立。
- 输出不构成法律、医疗或投资建议。
- 看起来笃定的表格仍是草稿；重要结论要回到原始材料核实。

### 隐私

把文本粘贴进聊天窗口可能被存储或传输。除非获准，不要粘贴机密、个人或受许可限制的材料。姓名、联系方式以及任何不能分享的内容应先脱敏。

文档化设置说明：来源已审阅；运行时未测试。

## 如何检查输出

1. 每一行是否都指向你确实粘贴过的摘录？
2. 支持性引文是否简短，且确实出现在那段摘录中？
3. 作者的主张是否与摘录描述的证据分开？
4. 日期和定义是否被保留，而不是被悄悄改写？
5. 是否标出了同源来源？
6. 三个追问是否真的是问题，而不是结论？
7. 任何假设是否被标注为基于你文本的推断？

## 常见问题

### 如果模型加进了自己记忆里的事实怎么办？

发纠正信息：只使用粘贴的摘录，其余标记为无依据。提示词本身已禁止浏览和凭记忆作答，但模型可能漂移；用一句简短提醒重新锚定，并要求它删除无依据的行。

### 一个论点被多段摘录重复提到，可以当作已确认吗？

提示词明确规定：在摘录间重复出现并不自动构成独立确认。把一致性当作更强证据之前，先检查这些摘录是否都依赖同一原始公告或数据集。

### 三个追问问题有什么用？

它们指向最关键的空白，并提示什么样的原始证据会有帮助。它们是研究建议，不是研究结论，也不是一份待办行动清单。

### 如果摘录之间互相矛盾怎么办？

比较部分应给出共识点、分歧点和缺失信息。如果摘录使用不同的度量或人群，应得到关于比较为何可能无效的解释，而不是被硬选出一个胜者。

## 来源与致谢

本提示词为 TokRepo 原创提示词，采用 CC BY 4.0 许可。关于通用 AI 工作空间能力的参考背景来自所链接的提供商文档，标注审阅日期为 2026-10-04；该参考仅作背景，不意味着本提示词依赖任何具体功能。外部参考材料保留其自身权利。感谢提供商文档维护者提供的公开参考。

## 完整可复制提示词

充当面向非专业读者的证据整理者。我会粘贴两个或更多来源摘录，并在可获得时附上它们的标题、URL 和日期，以及一个具体问题。仅使用这些摘录。不要浏览、编造引文，或把记忆中的事实当作证据使用。如果问题或摘录缺失，请向我索要并停止。

构建一个主张表，包含这些列：主张、哪条摘录支持它、精确的简短支持性引文、来源锚点、证据实际确立了什么，以及仍未知什么。将作者的断言与摘录中描述的证据区分开。一个主张在多条摘录中重复出现，并不自动构成独立确认。注明来源似乎依赖同一原始公告或数据集的情况。

对一致点、分歧点和缺失信息作简短比较。保留日期和定义。如果摘录使用不同的衡量标准或人群，解释为什么表面上的比较可能无效。不要仅凭声望给来源排名，不要从相关性推断因果，也不要把证据缺失转化为不存在证据。

最后提出三个聚焦的后续问题，以解决最重要的缺口，并说明哪种一手证据会有帮助。这些是研究建议，不是发现。只引用摘录中的小部分；其余用原创语言概括。将任何假设清楚标注为基于所提供文本的推断。

这是一种组织辅助，不是法律、医疗或投资建议。不要联系他人、发布结论或更改任何记录。我必须根据原始材料确认重要结论。

问题和摘录：将它们粘贴到这里，并附上实际已知的来源标签。

## 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/evidence-comparison-prompt-claims-vs-sources-ec1f70ee
Author: Prompt Lab