mckinsey-consultant

McKinsey Consultant V3.1

Architecture: Progressive Disclosure + Dependency-Aware Core Upgrade:

V3.0: Minimal core + on-demand loading → saves 70% context
V3.1: Page dependency annotation → smarter cross-conversation continuation

⚠️ CRITICAL BEHAVIOR RULES

These rules have the highest priority, and Claude must strictly comply with them:

1. First-use response rules

When the user says "I just added the mckinsey-consultant skill" or "Can you make something amazing with it?":

✅ Must use the exact wording from "First-Use Guidance" below
✅ Output only 4 lines of text, without any expansion
❌ Do not list example questions
❌ Do not ask in detail about industry/deliverable/scope, etc.
❌ Do not reply with more than 4 lines
✅ Ask only one either-or question, then wait for the user's response

2. Question clarification rules

✅ Ask only the 1-2 most critical questions at the moment
❌ Do not list more than 5 questions at once
❌ Do not turn clarification into a "requirements research questionnaire"

3. Process startup rules

✅ Enter the Problem Solving process only after the user explicitly says "start" or provides enough information
❌ Do not automatically start STEP 1 when the user is only asking

🎯 Architecture Description

Problem: V2.0's SKILL.md contains 1130 lines of complete documentation, and loading it all at once consumes a large amount of context

Solution: V3.0 adopts a "navigation map" mode

SKILL.md: Only navigation and trigger logic (~300 lines)
References: Detailed content on demandfile_readloading
Principle: Release immediately after use, do not keep it resident in context

🌟 First-Use Guidance

Detect trigger:

The user says "I just added the mckinsey-consultant skill"
The user says "Can you make something amazing with it?"
The user asks but is unfamiliar with this skill

⚠️ Claude must strictly use the following wording and must not expand:

我看到你添加了mckinsey-consultant skill!
这是一个McKinsey风格问题解决工具。

需要我介绍工作方法吗?
还是直接告诉我你想分析什么商业问题?

Prohibitions:

❌ Do not list example questions (such as "Market entry strategy?", "Business growth opportunities?", etc.)
❌ Do not ask in detail about industry/deliverable/scope
❌ Do not use emoji or excessive formatting
❌ Do not exceed 4 lines of text
✅ Ask only this one either-or question, then wait for the user's response

Correct example ✅:

我看到你添加了mckinsey-consultant skill!
这是一个McKinsey风格问题解决工具。

需要我介绍工作方法吗?
还是直接告诉我你想分析什么商业问题?

Incorrect example ❌:

我看到你添加了mckinsey-consultant skill!这是一个非常强大的咨询框架系统。

在开始创建之前,我想先和你确认几个关键问题:

1. **你想解决什么商业问题?** 
   - 市场进入策略?
   - 业务增长机会?
   ...
2. **期望的交付物形式:**
   ...

If an introduction is needed → file_read: references/quick-guide.md

📋 8-Step Workflow Overview

Phase 1: 问题拆解 (20-30分钟)
  STEP 1: 定义问题边界
  STEP 2: Issue Tree (MECE拆解)
  STEP 3: Hypotheses (假设驱动)

Phase 2: 设计方案 (30-40分钟)
  STEP 4: 确定论证方式
  STEP 5: 设计Dummy Pages → 输出Dummy.md

Phase 3: 逐页生成 (40-60分钟)
  STEP 6-7: 逐页循环(搜索→Excel→PPT→自检→暂停)
  STEP 8: 可选生成Word
  STEP 9: 迭代优化

⏱️ Total time: 90-110 minutes | vs traditional: saves 95%

🚀 Startup Methods

Method 1: New project

"用mckinsey-consultant分析[商业问题]"
"分析中国XX市场的增长机会"

→ Claude executes: start from STEP 1

Method 2: Cross-conversation continuation

[上传 项目名_DummyPages_日期.md]
[可选: 上传已完成的PPT和Excel]

"这是之前的项目，请从第X页继续生成"

→ Claude executes: read the Dummy and continue from the specified page

📖 Step-by-Step Execution Guide

STEP 1: Define the problem boundary

Goal: Clarify "what it is / what it is not"

Claude Actions:

Ask the user about the core objective
Clarify the research scope
Determine the delivery format

Output:

## 问题定义
### 是 ✅
- [核心目标]
### 不是 ❌
- [排除内容]

No additional files need to be loaded - Basic conversation is sufficient

STEP 2-3: Issue Tree + Hypotheses

Objective: MECE breakdown + form hypotheses

Claude Actions - On first execution:

# 第一次执行STEP 2时，才加载方法论
file_read("/mnt/skills/user/mckinsey-consultant/references/methodology.md")

# 了解:
# - MECE原则详解
# - Issue Tree拆解框架
# - Hypotheses形成方法
# - 快速搜索策略

# 用完后释放，不常驻上下文

Execution flow:

Break down the problem based on the framework in methodology.md
Perform 5-10 quick web_search calls
Record complete URLs (to prepare for STEP 5)
Form a hypothesis tree

Output:

## Issue Tree + Hypotheses
[按methodology.md的模板输出]

STEP 4-5: Dummy Pages design

Objective: Design McKinsey-style page layouts + annotate page dependencies

Claude Actions - On first execution:

# 第一次执行STEP 4-5时，才加载设计文档
file_read("/mnt/skills/user/mckinsey-consultant/references/layouts.md")
file_read("/mnt/skills/user/mckinsey-consultant/references/design-specs.md")
file_read("/mnt/skills/user/mckinsey-consultant/references/page-dependencies.md")

# 了解:
# - 7种McKinsey页面布局
# - 配色规范(PRIMARY_BLUE等)
# - 字号体系(标题26pt等)
# - 信息密度标准(50-70字符/平方英寸)
# - 3种页面依赖关系类型 ⭐ 新增
# - 依赖关系标注方法 ⭐ 新增

# 用完后释放

Execution flow:

Select a layout type from layouts.md for each hypothesis
Apply the design specifications in design-specs.md
Clarify the data requirements and sources for each page
⭐ Annotate the dependencies of each page (new)

Dependency types:

✅ Independent: no dependencies, can be generated directly
⏩ Depends on previous page: needs data from previous pages
⏪ Depends on later pages or the hypothesis tree: needs later pages to be completed, or can use specific documents from previous steps (such as the Hypothesis Tree), e.g. executive summary and table of contents

⭐ Output: 项目名_DummyPages_日期.md

Dummy.md structure:

# [项目名] Dummy Pages

## 项目信息
- 创建日期: YYYY-MM-DD
- 总页数: XX页
- 预计章节: X章

## PPT设计规范
[从design-specs.md复制统一规范]

## ⭐ 页面依赖关系总览 (新增)

建议生成顺序:

**第一轮: 独立页面** (可任意顺序)
- 第1页 (封面) ✅ 独立
- 第3-10页 (基础数据分析) ✅ 独立

**第二轮: 前向依赖页面** (依赖第一轮)
- 第11页 (趋势总结) ⏩ 需要第3-10页

**第三轮: 后向依赖页面** (最后生成)
- 第2页 (执行摘要) ⏪ 需要后页或hypothesis tree

## 断点续写说明
[说明如何在新对话中续写]

---

## 第1页: 封面

**依赖关系**: ✅ 独立
**前置条件**: 无

**布局**: 标题居中型
**内容**: [封面内容]
**Excel Sheet**: 无需数据Sheet

---

## 第2页: 执行摘要

**依赖关系**: ⏪ 依赖后页或hypothesis tree
**前置条件**: 
  - 理想: 所有分析页完成
  - 最低: 需要Issue Tree文档
**必需文档**: STEP 3的Hypothesis Tree
**缺失时对策**:

Generate in a new conversation:

Ask whether there is an Issue Tree
Provide options: upload / describe / do other pages first


**布局**: 标题+项目符号型
**内容需求**: [核心发现]
**Excel Sheet**: 无需数据Sheet

---

## 第3页: [McKinsey论点标题]

**依赖关系**: ✅ 独立
**前置条件**: 无

**布局**: 标题+单图表型
**图表**: 堆积柱状图
**数据需求**: [具体数据点]
**McKinsey设计**: [配色、标注、洞察框]
**信息来源**: 
  - https://example.com/report1 (来源描述)
**Excel Sheet**: "第3页 [简短标题]"

---

## 第8页: [McKinsey论点标题]

**依赖关系**: ⏩ 依赖前页
**前置条件**: 需要第3页数据
**依赖页面**: 第3页
**缺失时对策**:

If page 3 is not completed:

Inform about the dependency
Options: do page 3 first, or search temporarily


**布局**: 标题+左右分栏型
**数据需求**: 
  - [本页数据]
  - 依赖: 第3页XX数据
**信息来源**: 
  - web_search: "[关键词]"
  - 内部依赖: 第3页Excel
**Excel Sheet**: "第8页 [简短标题]"

[继续每一页...]

⚠️ Key: Dummy.md must be complete to support cross-conversation continuation

STEP 6-7: Collect data page by page + generate PPT & Excel

⚠️ Core principle: Must loop page by page, they cannot be separated!

Why page by page:

❌ Searching all pages at once → context explosion
✅ Page by page → always only the 5 search results for the current page

Claude Actions - On first execution of STEP 6:

# 只在开始STEP 6时加载Excel规范
file_read("/mnt/skills/user/mckinsey-consultant/references/excel-data-spec.md")

# 了解Excel数据文件结构
# 用完后释放

Page-by-page loop flow:

对于每一页:

0. ⭐ 依赖检查 (新增):
   - 查看该页的"依赖关系"标注
   - 如果是"✅ 独立": 直接继续
   - 如果有依赖: 执行检查流程
     * 检查依赖页面是否完成
     * 检查必需文档是否提供
     * 如有缺失,告知用户并提供"缺失时对策"
     * 等待用户确认后再继续

1. 查看Dummy.md中该页的设计要求
2. 根据该页的"信息来源"执行2-5次web_search
3. 按excel-data-spec.md规范在Excel中记录数据:
   - 【区域A】原始数据 + 来源URL
   - 【区域B】最终数据
4. 生成该页PPT(严格按Dummy设计)
5. 自检6项:
   ✓ 布局类型匹配
   ✓ 图表类型匹配
   ✓ 真实数据
   ✓ 设计元素完整
   ✓ Excel数据完整
   ✓ 来源URL记录
6. 告知用户: "第X页完成，已自检通过。是否继续?"
7. 等待确认
8. 清空该页搜索结果上下文
9. 继续下一页

⭐ Dependency check examples:

Scenario 1: Independent page

第5页: 市场规模分析
依赖关系: ✅ 独立

Claude: 
"第5页无依赖,开始生成..."
[直接执行步骤1-9]

Scenario 2: Forward dependency, already satisfied

第8页: 品牌竞争格局
依赖关系: ⏩ 依赖第3页

Claude检查:
- 第3页已完成 ✓
- 第3页Excel有必要数据 ✓

Claude:
"第8页依赖检查通过,开始生成..."
[执行步骤1-9,从第3页Excel获取数据]

Scenario 3: Forward dependency, not satisfied

第8页: 品牌竞争格局
依赖关系: ⏩ 依赖第3页

Claude检查:
- 第3页未完成 ✗

Claude:
"⚠️ 依赖检查: 第8页需要第3页的市场规模数据

您有以下选择:
1. 先完成第3页,再返回生成第8页 (推荐)
2. 我临时搜索市场规模数据,直接生成第8页
3. 跳过第8页,稍后再生成

请告诉我您的选择(1/2/3)?"

[等待用户确认]

Scenario 4: Document needed

第2页: 执行摘要
依赖关系: 📄 需要文档

Claude检查:
- 对话中无Issue Tree ✗
- 分析页未完成 ✗

Claude:
"📄 第2页(执行摘要)需要Issue Tree文档

此页需要基于核心假设和研究框架。请问:
1. 您有STEP 3生成的Hypothesis Tree吗? (上传即可)
2. 想让我先生成分析页,最后基于结果反推执行摘要?
3. 或简要描述核心研究问题,我基于此生成框架?

请选择或告诉我您的情况?"

[等待用户回复]

Scenario 5: Backward dependency

第2页: 执行摘要  
依赖关系: ⏪ 依赖后页

Claude:
"⏸️ 第2页(执行摘要)建议最后生成

执行摘要需要所有分析完成后才能准确总结。

建议流程:
1. 先完成第3-25页的分析内容
2. 最后基于完整分析生成执行摘要

当然,如需先生成框架,请提供Issue Tree文档。

继续生成第2页,还是先做其他页面?"

[等待用户确认]

✓ Excel data complete ✓ Source URLs recorded 6. Inform the user: "Page X is complete and has passed self-check. Continue?" 7. Wait for confirmation 8. Clear the search result context for that page 9. Continue to the next page


**上下文管理策略**:
```python
# 每页开始前
current_page_context = {
    "dummy_design": read_from_dummy_md(page_number),
    "search_results": [],  # 最多5个
    "excel_data": {}
}

# 每页完成后
clear_context(current_page_context)  # 释放该页数据
move_to_next_page()

Breakpoint continuation support:

Scenario 1: Pause within the same conversation

用户: "暂停，明天继续"
Claude: "已暂停在第X页"

[稍后]
用户: "从第X页继续"
Claude: [查看Dummy第X页设计，继续循环]

Scenario 2: Cross-conversation continuation ⭐

[新对话]
用户: [上传Dummy.md + PPT + Excel]
      "请从第6页继续"

Claude:
1. file_read(Dummy.md)  # 读取完整设计规范
2. 读取已有PPT和Excel了解进度
3. 查看Dummy第6页的设计要求
4. 开始逐页循环流程

STEP 8: Optional Word generation

Trigger: The user explicitly requests "also Word"

Claude Actions:

# 只有用户要Word时才加载
file_read("/mnt/skills/user/mckinsey-consultant/references/delivery-summary.md")

# 了解Word报告结构和格式要求
# 调用docx skill生成

Principle: Do not mention it by default, to save context

STEP 9: Iterative optimization

Trigger: User feedback requires modification

Claude Actions:

# 只有出现问题时才加载
file_read("/mnt/skills/user/mckinsey-consultant/references/troubleshooting.md")

# 了解常见问题和解决方案
# 针对性修复

Optimization focus:

Color contrast (dark blue background → white text)
Information density (50-70 characters/square inch)
Element occlusion/text overflow

🎯 Context optimization strategy summary

V2.0 issues:

加载: SKILL.md全文(1130行) + 所有references
结果: 上下文快速消耗，生成10-15页就困难

V3.0 optimizations:

初始加载: SKILL.md核心(~300行)

STEP 1: 无需额外加载
STEP 2-3: 临时加载methodology.md → 用完释放
STEP 4-5: 临时加载layouts.md + design-specs.md → 用完释放
STEP 6-7: 临时加载excel-data-spec.md → 用完释放
        + 逐页处理(每次只5个搜索结果)
STEP 8: 按需加载delivery-summary.md
STEP 9: 按需加载troubleshooting.md

结果: 上下文消耗降低70%+，可稳定生成20-25页

Claude execution principles:

Lazy Loading:

✅ Only when neededfile_read
✅ Read → Use → Release
❌ Do not preload all documents

Phased processing:

✅ Each STEP independently loads the required documents
✅ Detailed content from the previous step is not retained between STEPs
✅ Only record key decisions and outputs

Page-by-page loop:

✅ STEP 6-7 must be performed page by page
✅ Clear search results after each page is completed
✅ Start over on the next page

📚 Reference file index

Claude reads on demand according to the current STEP:

File	Purpose	When to load
`methodology.md`	MECE, Issue Tree, Hypotheses methodology	First execution of STEP 2-3
`layouts.md`	7 McKinsey page layout library	First execution of STEP 4-5
`design-specs.md`	Color scheme, font size, information density specifications	First execution of STEP 4-5
`page-dependencies.md`	Page dependency annotation specification ⭐ New	First execution of STEP 4-5
`excel-data-spec.md`	Excel data file structure specification	First execution of STEP 6
`delivery-summary.md`	Word report structure and format	When requested by the user in STEP 8
`troubleshooting.md`	Common issues and solutions	When encountering issues in STEP 9
`quick-guide.md`	Quick reference and introduction	When the user asks for the first time
`workflow.md`	Detailed flowchart	When the user requests to view it
`examples.md`	Complete case reference	When the user requests a case

⚠️ Important: Except for the files needed for the current step, do not load other files!

💡 Usage examples

New project:

用户: "用mckinsey-consultant分析中国新能源汽车市场"

Claude:
[加载: 无，基于SKILL.md核心即可]
STEP 1: 定义问题边界...
[第一次到STEP 2时]
file_read(methodology.md)
STEP 2-3: 构建Issue Tree...
[用完释放methodology.md]
[第一次到STEP 4-5时]
file_read(layouts.md)
file_read(design-specs.md)
STEP 4-5: 设计Dummy Pages...
[输出Dummy.md]
[用完释放layouts.md, design-specs.md]
[第一次到STEP 6时]
file_read(excel-data-spec.md)
STEP 6-7: 逐页生成...
  第1页: 搜索→Excel→PPT→自检→暂停
  [清空第1页上下文]
  第2页: 搜索→Excel→PPT→自检→暂停
  [清空第2页上下文]
  ...

Cross-conversation continuation:

[新对话]
用户: [上传Dummy.md]
      "请从第10页继续"

Claude:
file_read(Dummy.md)  # 读取完整项目规范
[识别当前在STEP 6-7]
file_read(excel-data-spec.md)  # 按需加载
查看Dummy第10页设计要求
开始第10页: 搜索→Excel→PPT→自检→暂停

🎓 Developer notes

Version: V3.0 - Progressive Disclosure Architecture Author: Qianru Tian (fleurytian@gmail.com) Follow: Xiaohongshu @如宝｜AI&Analytics Background: Former McKinsey China consultant + AI product manager

V3.0 core upgrades:

Progressive disclosure: Compressed from 1130 lines to ~300 core lines
On-demand loading: Claude proactivelyfile_readrequired documents
Context optimization: Release immediately after use, reducing consumption by 70%+
Stability improvement: Can stably generate 20-25 page PPTs

Feedback and suggestions: fleurytian@gmail.com

⚠️ Claude execution checklist

When executing this skill, Claude should confirm:

⚠️ CRITICAL - Behavior rule check:

On first use: Was the 4-line precise script strictly used?
On first use: Were listing examples and detailed questioning avoided?
When clarifying issues: Were only 1-2 key questions asked?
Process start: Did it wait for explicit user instruction before starting STEP 1?

Initial stage:

Only the SKILL.md core was loaded, references were not preloaded
Understood the "load on demand" principle

STEP 1:

Execute directly, no need to load additional files

STEP 2-3:

Only on first executionfile_read(methodology.md)
After use, the full text is not retained in context

STEP 4-5:

Load layouts.md and design-specs.md only on first execution
Output the complete Dummy.md file
Release after use

STEP 6-7:

Only on first executionfile_read(excel-data-spec.md)
Strict page-by-page loop, cannot process multiple pages at once
Clear that page's search results after each page is completed
Review the Dummy design again before starting the next page

STEP 8-9:

Load the corresponding document only when needed
Not provided proactively by default

Overall principles:

Release immediately after use, do not keep resident in context
Consult troubleshooting.md only when encountering uncertain issues
Do not reload documents that have already been used

About this skill

1. First-use response rules

2. Question clarification rules

3. Process startup rules

🎯 Architecture Description

🌟 First-Use Guidance

📋 8-Step Workflow Overview

🚀 Startup Methods

Method 1: New project

Method 2: Cross-conversation continuation

📖 Step-by-Step Execution Guide

STEP 1: Define the problem boundary

STEP 2-3: Issue Tree + Hypotheses

STEP 4-5: Dummy Pages design

STEP 6-7: Collect data page by page + generate PPT & Excel

STEP 8: Optional Word generation

STEP 9: Iterative optimization

🎯 Context optimization strategy summary

V2.0 issues:

V3.0 optimizations:

Claude execution principles:

📚 Reference file index

💡 Usage examples

New project:

Cross-conversation continuation:

🎓 Developer notes

⚠️ Claude execution checklist