如何让 Claude 像博士生一样研究：斯坦福 STORM 方法

Nav Toor 2026-07-24 11.3 KB 3.1M 5.3K 𝕏 文章列表

文章介绍了斯坦福大学开发的 STORM 研究系统，该方法通过多视角提问和信息检索生成高质量文章。作者将其简化为 4 个提示词，让用户能在 5 分钟内利用 Claude 生成类似博士水平的深度研究报告，包含视角分析、矛盾映射、综合综述和自我评估。

Claude STORM 提示词研究方法 AI工具 Prompt Engineering 斯坦福高效工作 Agent 多视角分析

# The Stanford STORM Method: How to Make Claude Research Like a PhD in Minutes **作者**: Nav Toor **日期**: 2026-06-17T10:36:52.000Z **来源**: [https://x.com/heynavtoor/status/2067194761446920264](https://x.com/heynavtoor/status/2067194761446920264) --- ![Image](/doc/3394/image/img_692985a68cb7.jpg) Most people use Claude like a search box. Ask, answer, close tab. They are leaving the best feature locked. Save this :) Stanford built a research system called STORM. In peer reviewed testing it produced articles 25 percent more organized than the next best method. It is open source. It is free. Almost nobody knows you can run the same idea inside Claude with 4 prompts. No software. No GitHub. No setup. Just paste. 5 minutes from now you will know more about your topic than people who spent days reading. Here is the full method. ![Image](/doc/3394/image/img_8acd17f5be92.jpg) ## Phase 1: What STORM Actually Is STORM stands for Synthesis of Topic Outlines through Retrieval and Multi perspective Question Asking. It was published at NAACL 2024 by the Stanford OVAL Lab. You can try the live version at storm.genie.stanford.edu. Free. No sign up. Type a topic and watch it write a sourced article in front of you. A 12 minute walkthrough video is here: STORM by Stanford on YouTube. Worth watching once. The full code is at github.com/stanford-oval/storm. MIT license. Run it on your own laptop if you want. But here is the real prize. You do not need any of it. The Stanford method is just a way of thinking. You can run that same thinking inside Claude with 4 copy paste prompts. That is what the rest of this article is. ![Image](/doc/3394/image/img_e01e84aa1eff.jpg) ## Phase 2: Why One Prompt Will Always Fail When you ask Claude "tell me about X" you get the majority view. The most common framing. The surface. What you do not get is the practitioner who works with X every day. The skeptic who thinks the field is wrong. The economist who follows the money. The historian who has seen the pattern before. The academic who actually read the studies. Those five voices all see different things. That is what a PhD student does. They do not ask one question. They ask five. The Stanford paper proved this with numbers. Articles built from multiple perspectives were 25 percent more organized and 10 percent broader in coverage than articles built the normal way. That is the entire breakthrough. Multi perspective questioning catches blind spots that single prompt research never sees. A PhD level research job takes 40 to 60 hours of human reading. Most people cannot spare that. STORM compresses it. The four prompts below compress it further. Five minutes total. ## Phase 3: Prompt 1, The Multi Perspective Scan This is the heart of the method. Paste this into Claude. Replace the topic in line 1. ``` I need to research [YOUR TOPIC]. Simulate 5 different expert perspectives on this topic: 1. THE PRACTITIONER: works with this daily. What do they know that academics miss? What practical realities are usually ignored? 2. THE ACADEMIC: has studied this for years. What does the peer reviewed evidence actually say? Where does the evidence contradict popular belief? 3. THE SKEPTIC: thinks the mainstream view is wrong. What is the strongest counterargument? What evidence do proponents conveniently ignore? 4. THE ECONOMIST: follows the money. Who profits from the current narrative? What financial incentives shape the research? 5. THE HISTORIAN: has seen similar patterns before. What historical parallels exist? What can we learn from how those played out? For each perspective give me: - Their core position in 2 sentences - The strongest evidence supporting their view - The one thing they would tell me that no other perspective would ``` What comes back: five very different reads of the same topic. The practitioner sees what the academic misses. The skeptic challenges what the practitioner assumes. The economist exposes incentives the academic ignores. The historian provides patterns the economist cannot see. This is 60 seconds of work that catches what one prompt never finds. ## Phase 4: Prompt 2, The Contradiction Map Now make Claude find where the 5 voices fight. The fights are where real understanding lives. ``` Based on the 5 perspectives above, map the contradictions: 1. Where do two or more perspectives directly contradict each other? List each conflict with the specific claims that clash. 2. Which perspective has the strongest evidence? Which has the weakest? Why? 3. What is the one question that, if answered, would resolve the biggest contradiction? 4. What does EVERY perspective agree on? (This is likely true. Even opponents confirm it.) 5. What topic did NONE of the perspectives address? (This is the blind spot in the whole field. Often the most valuable finding.) ``` What comes back: a map of where experts disagree and why. Most people skip this step. It is the step that separates surface understanding from real expertise. > If all 5 perspectives agree, it is probably true. If nobody addressed a topic, you just found the gap in the entire field. ![Image](/doc/3394/image/img_2852e28592d2.jpg) ## Phase 5: Prompt 3, The Synthesis Now make Claude pull everything together into a research briefing. ``` Synthesize everything from the 5 perspectives and the contradiction map into a research briefing: 1. THE ONE PARAGRAPH SUMMARY: explain this topic as if briefing a CEO who has 60 seconds and needs nuance, not just the headline. 2. THE 5 KEY FINDINGS: most important things I now know, ranked by reliability. For each, note which perspectives support it and which challenge it. 3. THE HIDDEN CONNECTION: one non obvious link between findings that only shows up when you look at all 5 perspectives together. 4. THE ACTIONABLE INSIGHT: based on all the evidence, what should someone in [YOUR ROLE] actually DO differently? Be specific. 5. THE FRONTIER QUESTION: the one question that, if answered, would change everything about how we understand this topic. ``` What comes back: a briefing no single expert could write. It accounts for every angle, names the contradictions, ranks reliability, and lands on a specific action. This is what a PhD student would produce in 48 hours. You just got it in 90 seconds. ## Phase 6: Prompt 4, The Peer Review STORM has one known weakness. Stanford's own researchers flagged it. The system does not self critique. Source bias and fact misassociation sneak in. This prompt fixes that by making Claude grade its own work. ``` Now peer review your own research briefing: 1. CONFIDENCE SCORES: rate each of the 5 key findings on a 1 to 10 scale for reliability. Explain each score. 2. WEAKEST LINK: which claim are you least confident in? What specific info would you need to verify it? 3. BIAS CHECK: which perspective might be overrepresented in your synthesis? Did one voice dominate? 4. MISSING PERSPECTIVE: is there a 6th angle I should have included that would change the conclusions? 5. OVERALL GRADE: if a Stanford professor reviewed this briefing, what grade would they give and why? What would they tell me to fix? ``` What comes back: an honest read of your own research. Strong claims, weak claims, biases, missing angles. Real peer review takes months. You just did it in 60 seconds. ![Image](/doc/3394/image/img_9fe4972f97b3.jpg) ## Phase 7: The 5 Minute Workflow Minute 1: Prompt 1. You have 5 expert views. Minutes 2 to 3: Prompt 2. You have a contradiction map. Minutes 3 to 4: Prompt 3. You have a research briefing. Minute 5: Prompt 4. You know what is reliable and what is not. Total time: 5 minutes. Output: a multi perspective briefing with contradiction analysis, synthesis, a specific action, and a reliability score. A PhD student takes 40 to 60 hours to produce this by hand. Not because they are slow. Because reading from 5 angles, mapping contradictions, synthesizing, and self critiquing is genuinely a 40 hour job for one human brain. ![Image](/doc/3394/image/img_be2bdd8e9345.jpg) ## Phase 8: 7 Ways to Use This Starting Today 1. Before writing any article or report. Run the 4 prompts. Your piece will cover angles nobody else thought of. 2. Before a major business decision. Get all 5 perspectives. The practitioner tells you what works in reality. The skeptic tells you what could go wrong. The economist tells you who profits. 3. Before a job interview. Research the company from 5 angles in 5 minutes. The practitioner view gives you insider language. The skeptic view gives you sharp questions. You walk in more prepared than anyone in the room. 4. Before investing. Bull case, bear case, historical parallel, incentive map, academic evidence. In 5 minutes. The contradiction map shows where the actual risk lives. 5. Before learning a new skill. Map the field from 5 angles. The practitioner tells you what to learn first. The academic tells you the theory. The skeptic tells you what is overhyped. You skip the noise. 6. Before a negotiation. Research the other side from 5 perspectives. Understand their incentives, weaknesses, historical behavior. You walk in with structural advantage. 7. Before any presentation. Run STORM on your topic. Your slides will answer objections before the audience raises them. Your Q and A will feel effortless. ## The Persona Block You are someone who reads. You ask hard questions. You do not want a 200 word summary that sounds smart but says nothing. You want to actually understand things. Quickly. With sources. The way a Stanford grad student would. Without the six year tuition bill. That is who this method is for. If that is you, save this article. Use the 4 prompts. Watch the difference. ![Image](/doc/3394/image/img_753123b1b0cf.jpg) ## The Uncomfortable Truth Here is what nobody is saying out loud. The Stanford team published this in 2024. The paper is peer reviewed. The code is open source. The live tool is free. The method is four prompts. And almost nobody uses it. We are in an 18 month window. The people who learn how to research with AI properly will out think the people who do not. By a lot. Not because they are smarter. Because they are running 5 perspectives, a contradiction map, a synthesis, and a peer review while everyone else is reading the first Google result. In 18 months, this kind of workflow will be baked into every tool. The edge will be gone. Today it is still a secret hiding in plain sight. Pick the topic you need to research most. Open Claude. Paste Prompt 1. Five minutes from now you will know more than people who spent days reading. The prompts are above. Stanford proved the method. The rest is up to you. hope this was useful. Nav ❤️ ## 相关链接 - [Nav Toor](https://x.com/heynavtoor) - [@heynavtoor](https://x.com/heynavtoor) - [storm.genie.stanford.edu](https://storm.genie.stanford.edu/) - [github.com/stanford-oval/storm](https://github.com/stanford-oval/storm) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [6:36 PM · Jun 17, 2026](https://x.com/heynavtoor/status/2067194761446920264) - [3M Views](https://x.com/heynavtoor/status/2067194761446920264/analytics) - [View quotes](https://x.com/heynavtoor/status/2067194761446920264/quotes) --- *导出时间: 2026/7/24 16:01:06* --- ## 中文翻译 # 斯坦福 STORM 方法：如何在几分钟内让像博士生一样进行深度研究 **作者**: Nav Toor **日期**: 2026-06-17T10:36:52.000Z **来源**: [https://x.com/heynavtoor/status/2067194761446920264](https://x.com/heynavtoor/status/2067194761446920264) --- ![Image](/doc/3394/image/img_692985a68cb7.jpg) 大多数人把 Claude 当作一个搜索框来用。提问，得到答案，关闭标签页。他们把最强大的功能锁住了。把这篇文章存起来吧 :) 斯坦福构建了一个名为 STORM 的研究系统。在同行评审测试中，它生成的文章比次优方法更有条理 25%。它是开源的。它是免费的。几乎没人知道你可以在 Claude 内部用 4 个提示词运行同样的思路。不需要软件。不需要 GitHub。不需要设置。只需粘贴。5 分钟后，你对你选题的了解将超过那些花了几天时间阅读的人。以下是完整的方法。 ![Image](/doc/3394/image/img_8acd17f5be92.jpg) ## 第 1 阶段：STORM 到底是什么 STORM 代表 Synthesis of Topic Outlines through Retrieval and Multi perspective Question Asking（通过检索和多视角提问综合主题大纲）。它由斯坦福 OVAL 实验室在 NAACL 2024 上发表。你可以在 storm.genie.stanford.edu 试用在线版本。免费。无需注册。输入一个主题，看着它在你面前写出一篇有来源的文章。这里有一个 12 分钟的演示视频：斯坦福的 STORM（YouTube）。值得看一遍。完整代码在 github.com/stanford-oval/storm。MIT 许可证。如果你愿意，可以在自己的笔记本电脑上运行。但真正的奖励在这里。你不需要上面提到的任何东西。斯坦福方法只是一种思维方式。你可以在 Claude 内部用 4 个复制粘贴的提示词运行同样的思维。这正是本文接下来的内容。 ![Image](/doc/3394/image/img_e01e84aa1eff.jpg) ## 第 2 阶段：为什么单个提示词总是会失败当你让 Claude “告诉我关于 X 的事”时，你得到的是主流观点。最常见的框架。表面现象。你得不到的是每天与 X 打交道的从业者。认为该领域观点错误的怀疑论者。追踪资金流向的经济学家。以前见过类似模式的历史学家。真正读过研究的学者。这五种声音看到的都是不同的东西。这正是博士生所做的工作。他们不只问一个问题。他们会问五个问题。斯坦福的论文用数据证明了这一点。基于多视角构建的文章比以常规方式构建的文章更有条理 25%，覆盖范围广 10%。这就是整个突破点。多视角提问能捕捉到单提示词研究永远看不到的盲点。博士水平的研究工作需要人类阅读 40 到 60 个小时。大多数人抽不出这么多时间。STORM 压缩了这个过程。下面的四个提示词进一步压缩了它。总共 5 分钟。 ## 第 3 阶段：提示词 1，多视角扫描这是该方法的核心。把这个粘贴到 Claude 中。替换第 1 行中的主题。 ``` 我需要研究 [你的主题]。模拟关于该主题的 5 种不同专家视角： 1. 从业者：每天与此打交道。他们知道哪些学术界忽略的东西？哪些实际现实通常被忽视？ 2. 学者：研究此课题多年。同行评审的证据实际上说明了什么？证据在哪里与流行观点相矛盾？ 3. 怀疑论者：认为主流观点是错的。最强的反驳论点是什么？支持者 conveniently 忽略了什么证据？ 4. 经济学家：追踪资金流向。谁从当前的叙事中获利？什么经济激励在塑造研究？ 5. 历史学家：以前见过类似的模式。存在哪些历史平行案例？我们能从那些案例的结局中学到什么？对于每个视角，请给我： - 他们的核心立场（2 句话） - 支持其观点的最有力证据 - 只有这个视角会告诉我的那一件事 ``` 你将得到：对同一主题的五种截然不同的解读。从业者看到了学者错过的东西。怀疑论者挑战从业者的假设。经济学家揭露了学者忽略的激励。历史学家提供了经济学家看不到的模式。这是 60 秒的工作，却能捕捉到单个提示词永远找不到的东西。 ## 第 4 阶段：提示词 2，矛盾映射图现在让 Claude 找出这 5 种声音在何处发生冲突。冲突是真正理解所在之处。 ``` 基于上述 5 个视角，绘制矛盾图： 1. 哪里有两个或更多视角直接相互矛盾？列出每个冲突以及具体冲突的主张。 2. 哪个视角的证据最强？哪个最弱？为什么？ 3. 如果有一个问题能得到答案，就能解决最大的矛盾，这个问题是什么？ 4. 所有视角都同意什么？（这很可能是真的。即使是反对者也会证实这一点。） 5. 没有任何视角涉及的主题是什么？（这是整个领域的盲点。通常是最有价值的发现。） ``` 你将得到：一张专家在何处不同意以及为何不同意的地图。大多数人跳过这一步。正是这一步将浅层理解与真正的专业知识区分开来。 > 如果所有 5 个视角都同意，那很可能是真的。如果没人涉及某个主题，你刚刚发现了整个领域的空白。 ![Image](/doc/3394/image/img_2852e28592d2.jpg) ## 第 5 阶段：提示词 3，综合现在让 Claude 将所有内容汇总成一份研究简报。 ``` 将 5 个视角和矛盾图中的所有内容综合成一份研究简报： 1. 一段式摘要：解释这个主题，就像向一位只有 60 秒时间的 CEO 做简报一样，他需要的是细微差别，而不仅仅是标题。 2. 5 个关键发现：我现在知道的最重要的事情，按可靠性排名。对于每一个，注明哪些视角支持它，哪些视角挑战它。 3. 隐藏的连接：一个非显而易见的发现之间的联系，只有当你同时审视所有 5 个视角时才会显现。 4. 可操作的洞察：基于所有证据，[你的角色] 中的人实际上应该做哪些不同的事情？要具体。 5. 前沿问题：如果有一个问题能得到答案，就会改变我们要对该主题的所有理解，这个问题是什么？ ``` 你将得到：一份没有任何单一专家能写出的简报。它顾及了每个角度，指出了矛盾，对可靠性进行了排名，并得出了具体的行动。这就是博士生在 48 小时内会产出的成果。你在 90 秒内就拿到了。 ## 第 6 阶段：提示词 4，同行评审 STORM 有一个已知的弱点。斯坦福自己的研究人员指出了这一点。该系统不会自我批评。来源偏差和事实错误关联会悄悄混入。这个提示词通过让 Claude 给自己的工作打分来解决这个问题。 ``` 现在对你自己的研究简报进行同行评审： 1. 置信度评分：在 1 到 10 的量表上，对 5 个关键发现中的每一个的可靠性进行评分。解释每一个分数。 2. 最薄弱环节：你对哪个主张最不自信？你需要什么具体信息来验证它？ 3. 偏见检查：在你的综合中，哪个视角可能被过度代表了？是否有一种声音占据了主导地位？ 4. 缺失的视角：是否有第 6 个我本该包含的角度，它可能会改变结论？ 5. 总体评分：如果一位斯坦福教授评审这份简报，他们会给什么分数，为什么？他们会让我修复什么？ ``` 你将得到：对你自己研究的诚实评估。强有力的主张，薄弱的主张，偏见，缺失的角度。真正的同行评审需要数月时间。你在 60 秒内就完成了。 ![Image](/doc/3394/image/img_9fe4972f97b3.jpg) ## 第 7 阶段：5 分钟工作流第 1 分钟：提示词 1。你拥有了 5 个专家视角。第 2 到 3 分钟：提示词 2。你拥有一张矛盾图。第 3 到 4 分钟：提示词 3。你拥有一份研究简报。第 5 分钟：提示词 4。你知道什么是可靠的，什么不可靠。总时间：5 分钟。输出：一份包含矛盾分析、综合、具体行动和可靠性评分的多视角简报。一名博士生需要 40 到 60 小时才能手工产出这些。不是因为他们慢。因为从 5 个角度阅读、映射矛盾、综合和自我批评，对于一个人脑来说确实是一份 40 小时的工作。 ![Image](/doc/3394/image/img_be2bdd8e9345.jpg) ## 第 8 阶段：从今天开始使用此方法的 7 种方式 1. 在撰写任何文章或报告之前。运行这 4 个提示词。你的作品将涵盖其他人没有想到的角度。 2. 在做出重大商业决策之前。获取所有 5 个视角。从业者告诉你现实中什么有效。怀疑论者告诉你什么可能出错。经济学家告诉你谁获利。 3. 在求职面试之前。在 5 分钟内从 5 个角度研究公司。从业者视角给你内部行话。怀疑论者视角给你尖锐的问题。你走进房间时比房间里的任何人都准备充分。 4. 在投资之前。多头案例，空头案例，历史平行案例，激励图，学术证据。在 5 分钟内完成。矛盾图显示了实际风险所在。 5. 在学习一项新技能之前。从 5 个角度绘制该领域的地图。从业者告诉你先学什么。学者告诉你理论。怀疑论者告诉你什么是过度炒作的。你跳过噪音。 6. 在谈判之前。从 5 个视角研究对方。了解他们的激励、弱点、历史行为。你带着结构性优势走进去。 7. 在任何演示之前。对你的主题运行 STORM。你的幻灯片将在观众提出异议之前就回答它们。你的问答环节将显得毫不费力。 ## 人物画像你是一个阅读者。你问尖锐的问题。你不想要一份听起来很聪明但什么也没说的 200 字摘要。你想要真正理解事物。快速地。有来源地。就像斯坦福研究生那样。而没有六年的学费账单。这就是这个方法的适用对象。如果这就是你，请保存这篇文章。使用这 4 个提示词。观察其中的差异。 ![Image](/doc/3394/image/img_753123b1b0cf.jpg) ## 令人不安的真相这是没有人公开说出来的事情。斯坦福团队在 2024 年发表了这项成果。该论文经过同行评审。代码是开源的。在线工具是免费的。该方法只需要四个提示词。而且几乎没人使用它。我们处于一个 18 个月的窗口期。学会如何正确利用 AI 进行研究的人，将在思维能力上远超那些不会的人。差距会很大。不是因为他们更聪明。因为他们运行了 5 个视角、一张矛盾图、一个综合和一次同行评审，而其他人还在阅读谷歌搜索的第一个结果。 18 个月后，这种工作流将被内置到每个工具中。优势将消失。今天它仍然是一个隐藏在众目睽睽之下的秘密。选择你最需要研究的主题。打开 Claude。粘贴提示词 1。 5 分钟后，你将知道得比那些花了几天时间阅读的人更多。提示词就在上面。斯坦福证明了该方法。剩下的取决于你。希望这对你有用。Nav ❤️ ## 相关链接 - [Nav Toor](https://x.com/heynavtoor) - [@heynavtoor](https://x.com/heynavtoor) - [storm.genie.stanford.edu](https://storm.genie.stanford.edu/) - [github.com/stanford-oval/storm](https://github.com/stanford-oval/storm) - [升级到 Premium](https://x.com/i/premium_sign_up) - [6:36 PM · Jun 17, 2026](https://x.com/heynavtoor/status/2067194761446920264) - [3M 观看](https://x.com/heynavtoor/status/2067194761446920264/analytics) - [查看引用](https://x.com/heynavtoor/status/2067194761446920264/quotes) --- *导出时间: 2026/7/24 16:01:06*