pixiesgeosun STORM Research Method
Stanford's STORM framework adapted as a 4-stage sequential pipeline. Published at NAACL 2024 by Stanford OVAL Lab.
Critical rule: stages must run sequentially. Each stage produces full-length output. The next stage receives the complete accumulated results as context. Never compress multiple stages into a single model call — that kills depth.
Output Destination
/Users/pixies/Library/Mobile Documents/iCloud~md~obsidian/Documents/Mydoc/Inbox/STORM-[主题]-[YYYY-MM-DD].md
Execution Protocol
Step 1: Get topic from user
Ask for the research topic. If the user provides a role (e.g. "我是出版人"), note it for Stage 3's actionable insight.
Step 2: Run Stage 1 — Multi-Perspective Scan
Load references/prompts.md. Send Prompt 1 to the model with the user's topic.
Collect the FULL output. Do not summarize or truncate.
Expected output: 5 perpectives (Practitioner, Scholar, Skeptic, Economist, Historian), each containing: 2-sentence core stance, strongest evidence, unique insight.
Step 3: Run Stage 2 — Contradiction Map
Send Prompt 2 to the model, preceded by the complete Stage 1 output. Format:
以下是第一阶段的多视角扫描结果：
[粘贴 Stage 1 完整输出]
---
基于上面的 5 个视角，绘制矛盾地图：
[Prompt 2 原文]
Collect the FULL output.
Expected output: direct contradictions with specific claims, evidence strength ranking with reasons, key question to resolve the biggest conflict, consensus among all 5, blind spots no one addressed.
Step 4: Run Stage 3 — Synthesis
Send Prompt 3 to the model, preceded by the complete Stage 1 + Stage 2 output. Use the user's role in the actionable insight section. Format:
以下是前两阶段的研究结果：
[粘贴 Stage 1 完整输出]
[粘贴 Stage 2 完整输出]
---
将以上 5 个视角和矛盾地图中的所有内容综合成一份研究简报：
[Prompt 3 原文，其中 [你的角色] 替换为用户角色]
Collect the FULL output.
Expected output: CEO summary, 5 key findings with supporter/skeptic per finding and reliability score (1-10), hidden connections, actionable insight, frontier question.
Step 5: Run Stage 4 — Peer Review
Send Prompt 4 to the model, preceded by the complete Stage 1 + Stage 2 + Stage 3 output.
Collect the FULL output.
Expected output: confidence score per finding with explanation, weakest claim and what's needed to verify, bias check with specific examples, missing 6th perspective, overall grade with what a Stanford professor would say.
Step 6: Assemble and Save
Merge all 4 stages into a single .md file with the structure below. Save to Vault. Notify user of path.
Output Structure
---
title: "STORM 研究：[主题]"
method: Stanford STORM
date: YYYY-MM-DD
---
# STORM 研究：[主题]
## 第一阶段：多视角扫描
### 1. 实践者
...
### 2. 学者
...
### 3. 怀疑论者
...
### 4. 经济学家
...
### 5. 历史学家
...
## 第二阶段：矛盾地图
...
## 第三阶段：研究简报
...
## 第四阶段：同行评审
...
Quality Checklist
After each stage, verify before proceeding:
- Each perspective has named entities (people, places, works, data)
- Arguments are specific, not generic platitudes
- Evidence is cited or described concretely
- Output feels like it was produced by someone with domain expertise
Extended Capabilities
Deeper Dive on One Perspective
If the user says "再深入 X 视角", load and use Prompt 5 from references/prompts.md.
Decision Support
If the user has a specific decision, after the 4 stages, load and use Prompt 6 from references/prompts.md. Triggered by keywords: "该不该"、"要不要"、"怎么选"、"决策"、"投资"、"谈判".
Reference
- Original paper: NAACL 2024, Stanford OVAL Lab
- Online tool: https://storm.genie.stanford.edu/
- Open source: https://github.com/stanford-oval/storm (MIT)