原文核对：2026-10-13 · 中文正文 · 英文提示词与演示
场景定位
使用 ChatGPT 研究论文或课程，让子代理分工阅读、收集背景，并形成带图解的 Markdown 报告。
- 难度
- 中级
- 时间跨度
- 30 分钟
适合用于
- 学习陌生概念的个人。
- 适合分工阅读、收集背景、绘图和书面综合的高密度材料。
- 把一次阅读整理成带引用和术语表的可复用 Markdown 报告。
相关工具
Mermaid 不适用时，生成用于说明而非精确表达的视觉素材。
起步提示词
英文原文 · 可复制后按实际材料修改
I want to learn a new concept from this research paper: [paper path or URL].
Please run this as a subagent workflow:
- Spawn one subagent to map the paper's problem statement, contribution, method, experiments, and limitations.
- Spawn one subagent to gather prerequisite context and explain the background terms I need.
- Spawn one subagent to inspect the figures, tables, notation, and any claims that need careful verification.
- Wait for all subagents, reconcile disagreements, and avoid overclaiming beyond the source material.
Final output:
- create `notes/[concept-name]-report.md`
- include an executive summary, glossary, paper walkthrough, concept map, method diagram, evidence table, caveats, and open questions
- use Markdown-native Mermaid diagrams where diagrams help
- use imagegen to generate illustrative, non-exact visual assets when a Markdown-native diagram is not enough
- cite paper sections, pages, figures, or tables whenever possible
Constraints:
- do not treat the paper as ground truth if the evidence is weak
- separate what the paper claims from your interpretation
- call out missing background, assumptions, and follow-up reading简介
学习高密度论文或课程，需要建立可用的理解：解决什么问题、方法怎样工作、证据是什么、依赖哪些假设，以及哪些部分仍待研究。
ChatGPT 可以自动收集背景，将复杂概念转成图解。也适合使用子代理：分别梳理论文结构、补充前置知识、检查图表和符号，再由主代理协调结果，生成以后可复习的报告。
最终产物应便于复习，例如 notes/concept-report.md 或其他格式文档，包含摘要、术语表、逐步讲解、图解、证据表、局限和开放问题，避免只留下临时聊天回答。
定义学习目标
说明概念和期望产物。聚焦的问题通常比宽泛摘要更有用。
例如：
我想理解这篇论文的核心思想、方法的运行方式、实验为何支持或不支持其论断，以及接下来应读什么。
这样的范围提供了明确任务：讲清概念，同时保留不确定性、引用论断来源，并区分论文主张与 ChatGPT 的解释。
贯穿示例：分析研究论文
假设你要学习一篇关于陌生模型架构的论文，希望通过报告快速理解概念，而不必先通读整篇论文。
合适的产物可以包含：
- notes/paper-report.md：主要说明。
- notes/figures/method-flow.mmd 或内嵌 Mermaid：方法流程。
- notes/figures/concept-map.mmd 或小型 SVG：前置概念之间的关系。
- 将论断对应到论文章节、页码、图或表的证据表。
- 后续阅读清单与未解决问题。
目标是让学习更系统，并留下可长期使用的成果。
让子代理分工
每个子代理应有明确范围和返回格式。需要时明确要求 ChatGPT 创建子代理；并非每次阅读都需要，但长篇或概念密集的论文适合并行探索。
论文分析可以这样分工：
- 论文结构：提取问题、贡献、方法、实验、局限和声称的结果。
- 前置背景：解释术语、相关概念和论文假定读者了解的先前工作。
- 符号与图表：讲解公式、算法、示意图、图和表。
- 审慎审阅：检查证据是否支持论断，列出限制、缺失基线或不清楚的假设。
主代理应等待各子代理完成，比较答案并解决矛盾，再综合成连贯报告。
有针对性地补充背景
遇到不熟悉的前置知识时，让 ChatGPT 从获准来源补充背景，例如本地笔记、参考文献文件夹、链接论文、已启用的网页搜索或已连接知识库。
学习内部概念时，可以用插件连接多个来源，建立知识库。
限定补充范围，说明可靠来源的标准，以及外部背景在报告中的用途：
- 在术语表中定义前置概念。
- 加入简短的“阅读前需要了解的背景”。
- 把后续阅读与论文自身的论断分开。
- 标出来自论文之外的论断。
为报告生成图解
图解常能快速检验理解。Markdown 报告应使用贴近原材料、便于修改的图。
常见选择包括：
- 概念图：展示前置概念及其关系。
- 方法流程图：追踪输入、转换、模型组件与输出。
- 实验图：关联数据集、指标、基线与报告的论断。
- 局限图：区分假设、失败模式和开放问题。
目标支持时优先使用 Mermaid，否则使用纳入版本管理的小型 SVG 或 PNG。只有需要示意性、非精确视觉素材，或 Markdown 原生图无法表达时，才使用 ChatGPT 默认提供的 imagegen 系统技能。
编写 Markdown 报告
报告应足够独立完整，便于日后回看。建议结构：
- 摘要。
- 阅读前需了解的知识。
- 关键术语与符号。
- 论文逐步讲解。
- 方法图解。
- 证据表。
- 论文未能证明的内容。
- 开放问题与后续阅读。
尽可能保留来源引用。PDF 应标注页码、章节、图或表；无法提取准确页码时，要说明并改用章节或标题引用。
把报告用于持续学习
首版报告只是起点。阅读后继续提问，并让 ChatGPT 更新文档。
可用的追问包括：
- 应先理解方法的哪一部分？
- 哪个最简单的玩具示例能说明核心思想？
- 哪张图最支撑论文论证？
- 哪个论断最薄弱或证据最少？
- 如果想实现它，下一步应读什么？
需要实验时，补充一个复现简化思想的小 notebook 或脚本，并从 Markdown 报告链接过去，让解释和实验保持关联。
提示词示例：
官方提示词
英文原文 · 可复制后按实际材料修改
Generate a script that reproduces a simple example from this paper.
The script should be self-contained and runnable with minimal dependencies.
There should be a clear output I can review, such as a csv, plot, or other artifact.
If there are code examples in the paper, use them as reference to write the script.可选技能
仅在符合目标产物时使用技能：
- $jupyter-notebook：可运行的玩具示例、图表或轻量复现。
- $imagegen：不要求精确技术表达的示意素材。
- $slides：学习完成后将报告转成演示文稿。
大多数论文分析报告宜默认使用 Markdown 原生图或简单 SVG，相比生成位图，更便于比较版本、审阅和随理解变化更新。
提示词示例
先创建报告大纲
官方提示词
英文原文 · 可复制后按实际材料修改
Before writing the full report, inspect [paper path] and propose the report outline.
Include:
- the core concept the paper is trying to explain
- which sections or figures are most important
- which background terms need definitions
- which diagrams would help
- which subagent tasks you would spawn before drafting
Stop after the outline and wait for confirmation before creating files.绘制概念图解
官方提示词
英文原文 · 可复制后按实际材料修改
Read `notes/[concept-name]-report.md` and add diagrams that make the concept easier to understand.
Use Markdown-native Mermaid diagrams when possible. If the report destination cannot render Mermaid, create small checked-in SVG files instead and link them from the report.
Add:
- one concept map for prerequisites and related ideas
- one method flow diagram for inputs, transformations, and outputs
- one evidence map connecting claims to paper figures, tables, or sections
Keep the diagrams faithful to the report. Do not add unverified claims.将报告转成学习计划
官方提示词
英文原文 · 可复制后按实际材料修改
Use `notes/[concept-name]-report.md` to create a study plan for the next two reading sessions.
Include:
- what I should understand first
- which paper sections to reread
- which equations, figures, or tables need extra attention
- one toy example or notebook idea if experimentation would help
- follow-up readings and questions to resolve
Update the report with a short "Next study loop" section.展开英文原文对照
Use ChatGPT to study material such as research papers or courses, split the reading across subagents, gather context, and produce a Markdown report with diagrams.
Introduction
Learning a new concept from a dense paper or course requires more than just summarization. The goal is to build a working mental model: what problem it addresses, what the method actually does, which evidence supports it, what assumptions it depends on, and which parts you still need to investigate.
ChatGPT is useful here because it can automate the context gathering, and can turn complicated concepts into helpful diagrams or illustrations. This use case is also a good fit for subagents: one thread can read the paper for structure, another can gather prerequisite context, another can inspect figures and notation, and the main thread can reconcile the results into a report you can review later.
For this use case, the final artifact should be something you can easily review: a Markdown file such as notes/concept-report.md, or a document of another format. It should include a summary, glossary, walkthrough, diagrams, evidence table, limitations, and open questions instead of ending with a transient chat answer.
Define the learning goal
Start by naming the concept and the output you want. A narrow question makes the report more useful than a broad summary.
For example:
I want to understand the main idea in this research paper, how the method works, why the experiments support or do not support the claim, and what I should read next.
That scope gives ChatGPT a concrete job. It should teach you the concept, but it should also preserve uncertainty, cite where claims came from, and separate the paper’s claims from its own interpretation.
Running example: research paper analysis
Suppose you want to learn about a paper about an unfamiliar model architecture. You want a report that lets you understand the concept at a glance, without having to read the whole paper.
A good result might look like this:
- notes/paper-report.md with the main explanation.
- notes/figures/method-flow.mmd or an inline Mermaid diagram for the method.
- notes/figures/concept-map.mmd or a small SVG that shows how the prerequisite ideas relate.
- An evidence table that maps claims to paper sections, pages, figures, or tables.
- A list of follow-up readings and unresolved questions.
The point is to make the learning process more systematic and to leave behind a durable artifact.
Split the work across subagents
Subagents work best when each one has a bounded job and a clear return format. Ask ChatGPT to spawn them explicitly; ChatGPT does not need to use subagents for every reading task, but parallel exploration helps when the paper is long or conceptually dense.
For a research paper, a practical split is:
- Paper map: Extract the problem statement, contribution, method, experiments, limitations, and claimed results.
- Prerequisite context: Explain background terms, related concepts, and any prior work the paper assumes.
- Notation and figures: Walk through equations, algorithms, diagrams, figures, and tables.
- Skeptical reviewer: Check whether the evidence supports the claims, list caveats, and identify missing baselines or unclear assumptions.
The main agent should wait for those subagents, compare their answers, and resolve contradictions. ChatGPT will then synthesize the results into a coherent report.
Gather additional context deliberately
When the paper assumes background you do not have, ask ChatGPT to gather context from approved sources. That might mean local notes, a bibliography folder, linked papers, web search if enabled, or a connected knowledge base.
If you’re learning about an internal concept, you can connect multiple sources with plugins to create a knowledge base.
Keep this step bounded. Tell ChatGPT what counts as a reliable source and what the final report should do with external context:
- Define prerequisite terms in a glossary.
- Add a short “background you need first” section.
- Link follow-up readings separately from the paper’s own claims.
- Mark claims that come from outside the paper.
Generate diagrams for the report
Diagrams are often the fastest way to check whether you really understand a concept. For a Markdown report, ask ChatGPT for diagrams that stay close to the source material and are easy to revise.
Good defaults include:
- A concept map that shows prerequisite ideas and how they connect.
- A method flow diagram that traces inputs, transformations, model components, and outputs.
- An experiment map that connects datasets, metrics, baselines, and reported claims.
- A limitations diagram that separates assumptions, failure modes, and open questions.
For Markdown-first reports, ask for Mermaid when the destination supports it, or a small checked-in SVG/PNG asset when it does not. Ask ChatGPT to use the imagegen system skill, which comes with ChatGPT by default, only when you need an illustrative, non-exact visual or something that doesn’t fit in a Markdown-native diagram.
Write the Markdown report
Ask ChatGPT to make the report self-contained enough that you can return to it later. A useful structure is:
- Executive summary.
- What to know before reading.
- Key terms and notation.
- Paper walkthrough.
- Method diagram.
- Evidence table.
- What the paper does not prove.
- Open questions and follow-up reading.
The report should include source references wherever possible. For a PDF, ask for page, section, figure, or table references. If ChatGPT cannot extract exact page references, it should say that and use section or heading references instead.
Use the report as a study loop
The first report is a starting point. After reading it, ask follow-up questions and have ChatGPT revise the artifact.
Useful follow-ups include:
- Which part of this method should I understand first?
- What is the simplest toy example that demonstrates the core idea?
- Which figure is doing the most work in the paper’s argument?
- Which claim is weakest or least supported?
- What should I read next if I want to implement this?
When the concept requires experimentation, ask ChatGPT to add a small notebook or script that recreates a toy version of the idea. Keep that scratch work linked from the Markdown report so the explanation and the experiment stay together.
Example prompt:
官方提示词
英文原文 · 可复制后按实际材料修改
Generate a script that reproduces a simple example from this paper.
The script should be self-contained and runnable with minimal dependencies.
There should be a clear output I can review, such as a csv, plot, or other artifact.
If there are code examples in the paper, use them as reference to write the script.Skills to consider
Use skills only when they match the artifact you want:
- $jupyter-notebook for toy examples, charts, or lightweight reproductions that should be runnable.
- $imagegen for illustrative visual assets that do not need to be exact technical diagrams.
- $slides when you want to turn the report into a presentation after the learning pass is done.
For most paper-analysis reports, Markdown-native diagrams or simple SVG files are better defaults than a generated bitmap. They are easier to diff, review, and update when your understanding changes.
Suggested prompts
Create the Report Outline First
官方提示词
英文原文 · 可复制后按实际材料修改
Before writing the full report, inspect [paper path] and propose the report outline.
Include:
- the core concept the paper is trying to explain
- which sections or figures are most important
- which background terms need definitions
- which diagrams would help
- which subagent tasks you would spawn before drafting
Stop after the outline and wait for confirmation before creating files.Build Diagrams for the Concept
官方提示词
英文原文 · 可复制后按实际材料修改
Read `notes/[concept-name]-report.md` and add diagrams that make the concept easier to understand.
Use Markdown-native Mermaid diagrams when possible. If the report destination cannot render Mermaid, create small checked-in SVG files instead and link them from the report.
Add:
- one concept map for prerequisites and related ideas
- one method flow diagram for inputs, transformations, and outputs
- one evidence map connecting claims to paper figures, tables, or sections
Keep the diagrams faithful to the report. Do not add unverified claims.Turn the Report Into a Study Plan
官方提示词
英文原文 · 可复制后按实际材料修改
Use `notes/[concept-name]-report.md` to create a study plan for the next two reading sessions.
Include:
- what I should understand first
- which paper sections to reread
- which equations, figures, or tables need extra attention
- one toy example or notebook idea if experimentation would help
- follow-up readings and questions to resolve
Update the report with a short "Next study loop" section.