Use 10 independent Sub-agents to each play the role of "advisor" as one of the world's strongest minds, conduct parallel analysis on the material provided by the user, then cross-aggregate to produce a structured report.
Lineup: Dan Sullivan, Musk, Zhang Xiaolong, MrBeast, Munger, Thiel, Jobs, Bezos, Zhang Yiming, Ren Zhengfei
Core value:When analyses are independent but conclusions converge, credibility is extremely high; when conclusions diverge, that is often the most valuable insight.
Process overview
| Stage | Name | Objective | Execution method |
|---|---|---|---|
| 1 | Converge requirements | Clarify what to analyze, why to analyze it, and from which perspectives | Dialogue |
| 2 | Prepare material | Organize the analysis material and confirm each agent's input | Dialogue confirmation |
| 3 | Parallel analysis | 10 Sub-agents each analyze independently and produce independent reports | Agent tools in parallel |
| 4 | Cross-aggregation | Find consensus, find disagreements, and extract actionable recommendations | Main dialogue |
| 5 | Output report | Structured report + optional: save as a file | Main dialogue |
Stage 1: Converge requirements
Objective: Clarify three things—what to analyze, why, and from which perspectives.
Information that must be confirmed
1. Analysis subject
- What does the user want to analyze? (review data for a period of time, a product strategy, a business direction, a decision...)
- Is there any ready-made material file? (If so, ask the user to provide the path or content)
2. Analysis purpose
- What does the user want to get from the analysis? (diagnose problems, validate direction, discover blind spots, make decisions...)
- Is there any core question they especially want answered?
3. Perspective selection
By default, 10 built-in perspectives are provided (see "Built-in perspective library" below). The user can:
- Use all 10 by default: suitable for comprehensive diagnosis ("the world's 10 smartest minds as your advisors at the same time")
- Choose 3-5 of them: suitable for focusing on specific dimensions
- Custom perspectives: the user specifies other thinkers/frameworks, and the AI temporarily builds a corpus through WebSearch
If the user is unsure what to choose, recommend based on the analysis subject:
| Analysis subject type | Recommended perspectives |
|---|---|
| Comprehensive personal/creator review | All 10 |
| Product strategy | Musk (first principles) + Zhang Xiaolong (product intuition) + Jobs (ultimate experience) + Bezos (customer obsession) |
| Content growth | MrBeast (growth) + Zhang Xiaolong (authenticity of demand) + Zhang Yiming (algorithmic thinking) + Ren Zhengfei (principle of concentrated force) |
| Business decisions | Munger (multiple mental models) + Thiel (monopoly thinking) + Bezos (flywheel) + Dan Sullivan (10x thinking) |
| Career direction | Dan Sullivan (10x vs 2x) + Musk (contrarian) + Thiel (secrets) + Ren Zhengfei (long-term hard work) |
| Competitive strategy | Ren Zhengfei (pressure principle + needle-tip strategy) + Thiel (monopoly) + Munger (circle of competence) |
| Strategic review | Ren Zhengfei (Huawei's Winter + self-criticism) + Munger (invert, always invert) + Musk (first principles) |
Criteria for completing convergence
- The analysis subject is clear (there is material or it can be described clearly)
- The analysis purpose is clear (at least one core question)
- The perspective has been confirmed
Stage 2: Prepare materials
Goal: Organize the materials provided by the user into input that each agent can use directly.
Material organization rules
- If the user provided a file path: Read the file contents and organize them into a text summary
- If the user described it in the conversation: Extract the key information and form a structured "analysis brief"
- If the material is too long(more than 5,000 words): Extract the core data and key facts, and compress them into a version each agent can digest
Input package for each Agent
The input each Sub-agent receives must include:
- Role setting: You are [thinker's name], analyze using the methodology of [framework name]
- Analysis material: [organized material]
- Core question: The question the user most wants answered
- Framework reference: The corresponding reference file path (the sub-agent reads the full original text itself)
- Output requirements：
⚠️ Hard rules regarding framework references:
- Do not pre-inject the full reference text into the main prompt——stuffing the corpus of all perspectives into the prompt will make it too long, causing the agent to lose or degrade
- Instead: in each agent's prompt, tell it to read the corresponding reference file itself
- Each agent reads only its own copy, keeping the burden manageable
- The reference file contains the thinker's direct quotations, specific cases, and analysis question templates——the agent must base its analysis on these
- Must use the core concepts and terminology of that framework
- Must give clear judgments and recommendations, do not be ambiguous
- Must include "the harshest sentence"——a diagnosis the user may not want to hear but must hear
- No word limit; write as much as needed
- Tone: in the thinker's first-person voice, speaking directly to the user
After confirmation, proceed to parallel analysis
Stage 3: Parallel analysis (core stage)
Execution method: Use the sub-agent tool of the current Agent platform, launch the selected perspectives in batches according to the available concurrency, and ensure each perspective is analyzed independently; do not exceed the platform's concurrency limit.
Preparation before launch: first create the output folder
Before launching the agent,you must first create the output folder：
mkdir -p "多视角分析/分析报告-[主题简称]"
Only start the agent after the folder has been created. Each agent's prompt will tell it to write the report to this folder.
Sub-agent launch template
⚠️ Core strategy: Each agent reads the corpus itself and writes the report itself
- The main conversation does not pre-read reference files (to avoid the prompt becoming too long and causing the agent to lose information)
- Each agent reads its own reference file itself
- After each agent finishes its analysis, it writes the report to the specified file itself
- The main agent is only responsible for finally reading the reports of all selected perspectives to do cross-summary
The reference paths in the table below are relative to this skill directory. The main agent should first resolve them to absolute paths based on the actual installation location, then pass them to the sub-agents; do not hardcode .claude/skills or the local user directory.
The corpus files and output files corresponding to each perspective:
| # | Perspective | reference file | output file |
|---|---|---|---|
| 1 | Dan Sullivan | reference/01-10x-growth-dan-sullivan.md |
多视角分析/分析报告-[主题]/01-DanSullivan.md |
| 2 | Musk | reference/02-first-principles-elon-musk.md |
多视角分析/分析报告-[主题]/02-马斯克.md |
| 3 | Zhang Xiaolong | reference/03-product-intuition-zhang-xiaolong.md |
多视角分析/分析报告-[主题]/03-张小龙.md |
| 4 | MrBeast | reference/04-creator-growth-mrbeast.md |
多视角分析/分析报告-[主题]/04-MrBeast.md |
| 5 | Munger | reference/05-mental-models-charlie-munger.md |
多视角分析/分析报告-[主题]/05-芒格.md |
| 6 | Thiel | reference/06-zero-to-one-peter-thiel.md |
多视角分析/分析报告-[主题]/06-蒂尔.md |
| 7 | Jobs | reference/07-product-vision-steve-jobs.md |
多视角分析/分析报告-[主题]/07-乔布斯.md |
| 8 | Bezos | reference/08-customer-obsession-jeff-bezos.md |
多视角分析/分析报告-[主题]/08-贝佐斯.md |
| 9 | Zhang Yiming | reference/09-algorithm-thinking-zhang-yiming.md |
多视角分析/分析报告-[主题]/09-张一鸣.md |
| 10 | Ren Zhengfei | reference/10-strategic-survival-ren-zhengfei.md |
多视角分析/分析报告-[主题]/10-任正非.md |
Each agent's prompt template:
你是 [思想家名字]。用你的 [框架名] 来分析以下素材，回答用户的核心问题。
## 第一步：读取你的分析框架（必须执行）
请先用 Read 工具读取以下文件，这是你的完整分析框架语料：
[对应的 reference 文件绝对路径]
读取后，严格基于该文件中的框架、原话、案例和分析问题来做分析。不可跳过此步骤，不可用自己的知识替代文件内容。
## 分析素材
[注入整理后的用户素材]
## 用户的核心问题
[用户最想回答的问题]
## 输出要求
1. 用你的框架的核心概念做深度分析，不要泛泛而谈
2. 必须引用 reference 文件中的原话来支撑你的分析（用 > 引用格式标注）
3. 给出明确的判断和建议
4. 用第一人称口吻，直接对用户说话，像真人顾问一样
5. 最后必须有一段"最刺耳的一句话"——你认为用户最需要听到但可能最不想听的诊断
6. 该写多少写多少，深度优先于篇幅控制
7. 请用中文输出
## 最后一步：保存报告（必须执行）
分析完成后，用 Write 工具将你的完整分析报告写入以下文件：
[对应的输出文件绝对路径]
不可跳过此步骤。写完后你的任务就结束了。
Key rules
- Create the folder before launching: Only start the agent after mkdir is complete
- Must all be launched in parallel: All selected agents are launched in the same message. If 5 perspectives are selected, 5 agents must be launched, with none omitted
- Each agent writes its own file: The main agent is not responsible for writing the reports for each perspective, only for writing the summary
- Each agent is completely independent: Do not let any agent see the output of other agents
- Do not preset conclusions: Do not hint in the prompt that "you should find XX problem"
Phase 4: Cross-summary
Goal: Extract high-value insights from the report files already saved for each perspective.
⚠️ Prerequisite for execution: After all agents are complete, use the Read tool to read the report files in the output folder one by one and confirm that they have all been written. Then perform cross-analysis based on the file contents.
Summary framework
1. Consensus findings (agreement by multiple parties = high credibility)
- Scan all reports and find points where 3 or more perspectives reach the same conclusion
- These are the most reliable diagnoses—independent analyses converge, indicating "hard facts"
2. Interesting disagreements (conflicting perspectives = high-value insights)
- Find points where different perspectives give contradictory recommendations
- Analyze the reasons for the disagreement: Is it different framework assumptions? Or are they seeing problems at different levels?
- Disagreements are often more enlightening than consensus—they expose the multifaceted nature of the problem
3. Exclusive insights (proposed by only one perspective)
- Did any perspective propose an important viewpoint that no one else mentioned?
- These may be overlooked blind spots
4. Actionable recommendations (converge into actions)
- Extract specific action recommendations from all reports
- Sort by "importance x executability"
- Converge into the 3-5 most critical next-step actions
Summary output format
# [分析主题] — 多视角诊断报告
> 分析日期：YYYY-MM-DD
> 视角：[使用了哪些视角]
> 素材：[分析了什么]
---
## 多人共识（高可信度）
[3-5 个多数视角一致的核心结论]
## 有意思的分歧
[视角之间的矛盾点及其原因分析]
## 独家洞察
[只有某个视角提出的重要观点]
## 最刺耳的话（每人一句）
| 视角 | 最刺耳的一句 |
|------|------------|
| ... | ... |
## 下一步行动（3-5 条）
[按优先级排序的具体行动建议]
---
## 附：各视角完整报告
[保留每个 agent 的原始输出，用户可展开阅读]
Stage 5: Save and output (executed automatically, no user request needed)
⚠️ This is a mandatory step. After the agent finishes running, it must be executed automatically. Do not wait for the user to request it.
Save process
Step 1: Create the output folder
In the 多视角分析/ directory, create a folder named after the analysis topic:
多视角分析/
分析报告-[主题简称]/
00-汇总报告.md
01-Dan-Sullivan.md
02-Elon-Musk.md
03-Naval-Ravikant.md
04-张小龙.md
05-MrBeast.md
Step 2: Confirm that each perspective report has been written
Each agent has already written its report into the folder on its own (completed in Stage 3). Use ls or Glob to confirm that all files are present. If any agent failed to write successfully, the main agent should supplement it from the content returned by the agent.
Step 3: The main agent writes the summary report
Based on the cross-analysis in Stage 4, write the summary report to 00-汇总报告.md。The main agent writes only this one file.
Step 4: Inform the user
After saving is complete, tell the user:
- Which directory the files are saved in
- How many reports there are in total
- Show the core content of the summary report in the conversation (consensus + disagreements + action recommendations)
Naming rules
- Folder name:
分析报告-[主题简称](e.g.,分析报告-职业决策、分析报告-4月策略） - If a folder for the same topic already exists, add the date after the topic:
分析报告-职业决策-0410 - If there are custom perspectives (corpus built via WebSearch), also save them to the reference directory for reuse
Built-in perspective library (10 of the world's strongest minds)
Perspective 1: Dan Sullivan — 10x growth
Reference file：reference/01-10x-growth-dan-sullivan.md
Core dimensions: 10x vs 2x filter, 20%/80% rule, 4C cycle (commitment → courage → capability → confidence), unique ability, Want vs Need
Best suited for: personal growth strategy, time allocation, directional trade-offs, career decisions
Perspective 2: Musk — First principles
Reference file：reference/02-first-principles-elon-musk.md
Core dimensions: first-principles assumption breakdown, five-step engineering algorithm, probability x importance, contrarian judgment, time allocation
Best suited for: product strategy, process optimization, resource allocation decisions
Perspective 3: Zhang Xiaolong — Product intuition
Reference file：reference/03-product-intuition-zhang-xiaolong.md
Core dimensions: authenticity of demand, restraint and not doing, use and leave, forest vs palace, KPI is a byproduct
Best suited for: product design, content strategy, system evaluation
Perspective 4: MrBeast — Creator growth
Reference file：reference/04-creator-growth-mrbeast.md
Core dimensions: title/click-through rate, topic breadth testing, content flywheel, output and iteration, independent product thinking
Best suited for: content growth, topic evaluation, dissemination strategy
Perspective 5: Munger — Multiple mental models
Reference file：reference/05-mental-models-charlie-munger.md
Core dimensions: interdisciplinary mental models, invert, circle of competence, psychology of misjudgment, Lollapalooza effect, checklist, patiently wait for the fat pitch, avoid stupidity
Best suited for: decision quality diagnosis, cognitive bias audit, long-term strategy
Perspective 6: Thiel — Zero to One / Monopoly thinking
Reference file：reference/06-zero-to-one-peter-thiel.md
Core dimensions: zero to one vs one to N, monopoly thinking, contrarian thinking/secrets, power law, definite optimism, last mover advantage, the importance of sales
Best suited for: Positioning and differentiation, competitive strategy, direction choices
Perspective 7: Jobs — Ultimate experience and narrative
Reference files：reference/07-product-vision-steve-jobs.md
Core dimensions: Ultimate simplicity, the product is the marketing, reality distortion field, connecting the dots, taste and aesthetics, focus and saying no, Think Different
Best suited for: Product experience, brand narrative, content quality
Perspective 8: Bezos — Customer obsession and the flywheel
Reference files：reference/08-customer-obsession-jeff-bezos.md
Core dimensions: Day 1 thinking, customer obsession / working backwards, flywheel effect, long-termism, high standards / six-page memo, one-way door vs two-way door, regret minimization
Best suited for: Growth strategy, user insights, long-term planning
Perspective 9: Zhang Yiming — Algorithmic thinking
Reference files：reference/09-algorithm-thinking-zhang-yiming.md
Core dimensions: Delayed gratification, algorithmic / information distribution efficiency, think like a machine, always be starting up, Context not Control, talent density, global vision
Best suited for: Personal growth, organizational efficiency, information distribution strategy
Perspective 10: Ren Zhengfei — Strategic survival philosophy
Reference files：reference/10-strategic-survival-ren-zhengfei.md
Core dimensions: Principle of pressure (focus force into one point), Huawei's winter (crisis awareness), dedication to strivers, gray-scale management, deepen the beach and lower the weir, needle-tip strategy (focus on the main channel), self-criticism
Best suited for: Strategic focus, crisis management, resource allocation, competitive strategy, long-term survival
Custom perspective
If the user wants to add a perspective outside the built-in library, follow the process in reference/_how-to-create-reference.md . That document contains the corpus file template, collection strategy, quality standards, and checklist.
Core principles
- Independence is the lifeline: Each agent must analyze completely independently and must not see the output of other agents. This is the entire value of multi-perspective analysis
- Must be parallel: All agents are launched in the same message; do not run them serially. Serial execution causes later agents to be influenced by earlier ones
- Depth first: Better to have each perspective write a 3,000-word in-depth report than 5 shallow 500-word summaries
- Must have sharpness: Each agent must provide "the most jarring sentence." Polite but useless analysis is worse than none at all
- The synthesis must have a point of view: Cross-synthesis is not a simple list; it must analyze why consensus formed and why disagreements arose
- User decision-making power: Analysis is a tool; the decision belongs to the user. The report gives recommendations at the end and does not make decisions