2024 年,斯坦福 OVAL 实验室 [0,11] 在 NAACL 2024 发表了题为 Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models 的论文 [1,39],介绍了一个名为 STORM 的系统——全称是 Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking [0,31]。它要解决的是长文写作里最重的那部分负担:动笔之前的研究、找角度、搭结构 [42]。围绕这套方法流传最广的两句话是”产出比最强方法有条理 25%“和”5 个专家角色、4 条提示词、5 分钟出简报” [0],这两句话都需要精确的上下文。下面先讲系统实际怎么运转,再逐条核对这些说法。
它到底在做什么:自动完成”写作前的研究”
写一篇维基百科级的长文,对资深作者也是挑战,最花时间的是写作前的准备阶段:研究主题、收集来源、形成大纲 [42]。STORM 把这一过程拆成两个阶段:先是”写作前”阶段,在线检索参考文献、生成大纲,并模拟 AI 智能体之间的对话、用有针对性的问题丰富发现;然后进入”写作”阶段,基于大纲和参考文献生成带引用的全文 [43]。论文由 Shao、Jiang、Kanell、Xu、Khattab、Lam 撰写 [39,45]。
它同时提供三层入口:开源代码放在 github.com/stanford-oval/storm [31],官方托管页面 storm.genie.stanford.edu 免费使用、无需注册 [0,46]。但官方定位是”研究预览""研究原型”,用于交互式知识策展 [11,13],论文团队也把它描述为”辅助”写作准备阶段的工具 [42]。
核心机制:把”提出好问题”变成系统工程
论文把自动化研究的核心问题定义为”自动提出好问题”——直接让语言模型提问,效果并不好 [31]。为此 STORM 采用两个策略:
一是视角引导的提问:给定主题后,系统先从相似主题的已有文章中自动发现不同视角,再用这些视角控制提问过程 [31]。
二是模拟对话:系统模拟一位维基百科写作者与一位主题专家之间的对话,对话以网络来源为 grounding,让模型不断更新对主题的理解、提出追问 [31]。
这里需要澄清素材的一个常见误解。素材把机制概括成”模拟五类专家视角:从业者、怀疑论者、经济学家、历史学家、学者” [0,36],但这套名单是素材作者的教学框架,不是系统的固定设计:论文里的视角是自动发现的 [31],不同教程给出的角色名单也不一致(另一份教程写的是领域专家、怀疑论者、从业者、记者、新人 [34])。提示词的数量同样不统一:素材说 4 条 [0],另一份教程拆成 5 条 [39]。这些差异说明”五角色四提示”是对同一思想的手工改写,而不是系统本身。后续迭代 Co-STORM 引入了协作话语协议和轮次管理,让人工在不同节点介入对话 [10,12],直接输出带引用的 PDF 报告 [45]。
“25%更有序”的精确口径
素材引用最多的数字是”在同行评审的测试中,产出文章比最强方法有条理 25%” [0]。论文里的实际口径窄得多:团队用近期高质量维基百科文章构建了 FreshWiki 评测集,请有经验的维基百科编辑做人工评估 [43];与”大纲驱动的检索增强基线”相比,更多 STORM 文章被判定为”有条理”,增幅为 25 个百分点(绝对增幅),覆盖面宽 10% [1]。也就是说,25% 是”被判定有条理的文章比例”的百分点提升,对比对象是论文自己设置的基线,而不是泛指的最强方法 [1]。
同一项专家反馈还指出了新挑战:来源偏差的传递和无关事实的过度关联 [1]——这正是素材所说”来源偏差与事实错配”的出处 [0]。同时,被调研的维基百科作者认可 STORM 能帮上准备阶段的忙,但机器生成的文本质量还达不到人工精编文章的水平 [42]。放在更大背景里,LLM 的人工评估方法本身仍是研究热点:评估结果的可复现性 [3] 和标注者视角差异 [5] 都还在讨论中,自动指标与人工判断的一致性在医学诊断文本生成这类任务上也被证明很低 [7]。
四步提示链:把方法搬进通用 LLM 的边界
素材的核心实操建议是:不装软件、不跑代码,把 STORM 的思想压成 4 条提示词依次喂给 Claude,5 分钟得到”多视角简报” [0]。这本质上是提示词工程里的”提示链”:把复杂任务拆成顺序执行的步骤,上一步的输出作为下一步的输入 [47,49,50]。素材的四步——多视角扫描、矛盾映射、综合简报、同行评审——每一环都承接上一环的输出 [0]。
这套改写的价值在于把”多视角”变成了一种任何人可执行的思维流程,但它与真正的 STORM 有两个关键落差。第一,STORM 的两个引擎——自动发现视角、检索支撑的模拟对话——都依赖联网检索和来源 grounding [31],写作阶段也带着引用生成 [43,46];纯提示词版本没有检索这一步,事实核对完全交给模型自己的记忆。第二,提示链的可靠性高度依赖各步骤之间输出的质量和格式,某一步输出模糊,后续步骤就可能连锁出错 [47]。
素材还有一处数字需要降级使用:“博士级研究需要 40 到 60 小时人工阅读” [0,33] 没有统一依据——关于博士生该读多少,研究者明确表示没有硬性规则,取决于主题和对”读”的定义 [28]。这个对比是修辞,不是经过验证的等价关系。
已知弱点与正确使用方式
STORM 被作者们指出的已知弱点,正是素材第四步提示词要补的洞:系统不会自我批判,来源偏差和事实错误关联会溜进来 [0,1]。更根本地,LLM 自身的来源可信度评级并不可靠——研究显示政治偏见越强的模型配置,评级准确率越低 [17],所以”来源偏差”并不只在 STORM 这一层出现。素材的延伸解读也承认:如果底层模型在第一步就产生幻觉,后续的多视角讨论可能放大错误;方法效果依赖提示词写得好不好;“历史学家”视角还受限于模型训练数据的时效 [0]。
因此更稳妥的用法,是把这套流程当作决策前的扫描器而不是结论本身:用它快速铺开视角、找出分歧与盲区,再对关键论断回到原始来源核验。素材列举的写作、投资、面试、学习等场景都适用 [0],但产出应当作”待核验的研究草案”,而不是”已证实的答案”。
原文更正:素材称论文证明产出文章’比最强方法有条理25%‘,实际口径是与大纲驱动的检索增强基线相比、被维基百科编辑判定为有条理的文章比例提高25个百分点(绝对增幅),对比对象并非泛指的’最强方法’。 [1]
参考来源
- 素材原文(见文首来源链接)
- Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) - ACL Anthology
- What Are We Measuring in NLG? A Meta-Analysis of Evaluation Trends 2020-2025
- Finding Replicable Human Evaluations via Stable Ranking Probability - ACL Anthology
- Medium
- Human Evaluation of Large Language Models: A Review …
- Human-Centered AI Evaluation: Best Practices for Accuracy & Inclusivity
- Development of a Human Evaluation Framework and Correlation with Automated Metrics for Natural Language Generation of Medical Diagnoses - PMC
- Stanford’s STORM Method: Research Without the Blind Spots · Mika Reyes
- How I used STORM, a free AI research tool
- Testing Stanford’s Storm and Co-Storm - Amazing AI Article Writer Nov 2024 (Longer AI Edited Video)
- Storm Genie (Stanford) - Stanford University
- Co-Storm: FREE AI TOOL by STANFORD can convert TOPICS to LONG ARTICLES!
- Storm Genie (Stanford) - Stanford University
- Stanford STORM Explained: AI That Writes and Curates …
- Stanford’s STORM AI outperforms Perplexity & Google Deep Research
- Credentialed education from Stanford faculty | Stanford Online
- Accuracy and Political Bias of News Source Credibility Ratings by Large Language Models
- Paper Digest: NAACL 2024 Papers & Highlights – Resources | Paper Digest
- On the nature of real and perceived bias in the mainstream media
- Source Bias - Misinformation, Disinformation, and Propaganda - LibGuides at Cornell University
- The presence of unexpected biases in online fact-checking | HKS Misinformation Review
- The Truth About Online Media Bias and Accuracy | Fractl
- 评估人工智能在科学研究任务中的能力 | OpenAI
- 2025年人工智能指数报告
- Time for PhD supervision
- Doctoral Studies Policies | NYU Steinhardt
- Get a Life, PhD: What’s the matter with a forty-hour work week for academics?
- how much should doctoral researchers read? – patter
- AEA > Education & Programs > University Programs
- Echelon Scholars - PhD-Level Research for High Schoolers
- STORM: Synthesis of Topic Outlines through Retrieval and …
- Stanford STORM Method: 4 Prompts to Improve Writing with Claude
- Stanford’s STORM Method: Research Without the Blind Spots
- How to Build a Multi-Perspective AI Research Workflow Using the STORM Method | MindStudio
- Stanford’s STORM Method: Research Without the Blind Spots
- Nav Toor on X: “stop asking Claude one question and thinking you understand the topic. you don’t. Stanford proved a better way. it’s called STORM. peer reviewed. 25% more organized output. open source. the trick: don’t ask one question. ask five. from five different experts. >the https://t.co/SBFM8mkBbU” / X
- How to Build a Multi-Perspective AI Research Workflow Using the …
- Stanford Quietly Built a Research System 25% Better Than Humans. You Can Run It in Claude in 5 Minutes.
- [3.9K views · 19 reactions | Stanford’s OVAL lab built a research method called STORM that writes Wikipedia-grade, fully-cited articles from scratch — and almost nobody realizes you can run the exact same method inside Claude. In this video I turn STORM into 5 copy-paste prompts that make Claude research like a PhD instead of guessing like a chatbot.
🟦 Comment “STORM” and I’ll send you the free Prompt Pack (all 5 prompts, formatted + a worked example).
The 5 prompts (Stanford’s STORM pipeline, by hand):
- Perspective Discovery — map the 6–8 expert lenses a topic deserves
- Simulated Expert Interview — a cited Q&A per perspective (“cite a source or say unverified”)
- Curate & Outline — group, dedupe, flag contradictions, map every section to its sources
- Grounded Writing — section by section, every claim cited, no invented facts
- The Moderator Pass (Co-STORM) — surface the unknown unknowns + red-team source bias and red herrings
Why STORM beats “AI, write me a report”: it forces breadth (many perspectives ask things one prompt never would) and grounding (every answer tied to a real source) — so you get organized, broad, reference-backed research instead of a confident but shallow guess.
Method credit — all Stanford OVAL: • STORM paper — “Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models” (Shao, Jiang, Kanell, Xu, Khattab, Lam), NAACL 2024 • Open source: https://github.com/stanford-oval/storm • Free demo: https://storm.genie.stanford.edu This guide adapts their method into manual Claude prompts; it is not affiliated with Stanford.
Timestamps: 0:00 The method Stanford built 0:47 The problem with “write me a report” 1:39 What STORM actually is 2:53 The big idea: perspectives + grounding 3:51 Prompt 1 — Perspective Discovery 4:51 Prompt 2 — Simulated Expert Interview 5:51 Prompt 3 — Curate & Outline 6:48 Prompt 4 — Grounded Writing 7:39 Prompt 5 — Moderator + Red-Team (Co-STORM) 8:44 Putting it together 9:38 Free prompt pack + recap
📌 Follow Hyperautomation Labs everywhere: ▶️ YouTube: https://www.youtube.com/@hyperautomationlabs1045 📸 Instagram: https://www.instagram.com/hyperautomationlabs 👍 Facebook: https://www.facebook.com/HyperautomationLabs
📚 Go deeper — my full guides: https://hyperautomationlabs.gumroad.com (Claude Code Guide · OpenAI Codex Guide · Claude Cowork Sales · CCA-F Prep Kit)
#claude #ai #stormmethod #airesearch #claudeprompts #deepresearch #stanford | Hyperautomation Labs](https://www.facebook.com/hyperautomationlabs/videos/stanfords-secret-research-method-makes-claude-write-like-a-phd-5-prompts/1027640726406220) 40. 研究室人員可以試試看 STORM 是由史丹佛開放虛擬助理實驗室開發的 … 41. Medium 42. Stanford AI experiment “STORM” generates Wikipedia-style articles 43. Stanford University launches STORM, a new AI research tool that enables anyone to create Wikipedia-style reports on any topic - Tech Startups 44. This Stanford AI Tool Writes Wikipedia-Style Reports Instantly! | Co-STORM Tutorial 45. Co-STORM AI generates fact-checked research reports | ETIH EdTech News — EdTech Innovation Hub 46. STORM Stanford Review 2026: Free AI That Writes Wikipedia Articles 47. 第 1 章:提示链(Prompt Chaining) | Jimmy Song 48. 链式提示 | Prompt Engineering Guide 49. Prompt Chaining Langchain | IBM 50. 充分释放大模型的潜力——提示词工程的 16 种方式 | 亚马逊AWS官方博客 51. Prompt Engineering 教學:系統化提示設計完全指南 | 超智諮詢