Mck Speech Design Skill

McKinsey Speech Design Skill

Generate tailored, persona-fitted speech scripts from user-provided materials through a structured briefing and drafting workflow. When the source material is a PPTX file, automatically write the script back into each slide's speaker notes.

Overview

This skill transforms raw materials (slide decks, briefs, notes, strategy docs) into polished, ready-to-deliver speech scripts. The key differentiator: every speech is fitted to the speaker's specific persona, audience relationship, and event context — not generic.

The workflow has 3 stages (+ an automatic 4th stage for PPTX input):

1. Briefing: Gather context about speaker, audience, relationship, event, objectives, language, and speaking style reference
2. Architecture: Design the speech structure, timing, and narrative arc (McKinsey Pyramid Principle)
3. Drafting: Write the full script with per-section talking points, transitions, and speaker notes — informed by the chosen style reference
4. PPTX Injection (auto-triggered when input is .pptx): Write the script back into the PPTX file's speaker notes

Dependencies

This skill requires the following Python packages:

Install: pip install python-pptx python-docx

Note: markitdown is optional (used for PPTX text extraction but has a fallback).

Stage 1: Briefing

Collect context through conversation before writing anything. Do NOT start drafting until sufficient context is gathered.

Detect Input Type

First, check what the user has provided:

• If PPTX file (.pptx):
  1. First try: python -m markitdown <file.pptx> to extract text content.
  2. If markitdown fails or returns empty: Fall back to python-pptx extraction:

     from pptx import Presentation
     prs = Presentation("file.pptx")
     for i, slide in enumerate(prs.slides, 1):
         texts = []
         for shape in slide.shapes:
             if shape.has_text_frame:
                 texts.append(shape.text_frame.text)
         print(f"--- Slide {i} ---")
         print('\n'.join(texts))
     

  3. Also note the total slide count. Set input_type = "pptx" — this triggers Stage 4 later.
• If images/screenshots: Analyze the visual content directly.
• If documents (.docx, .pdf, .md): Read the content with appropriate tools.

Required Information (collect in priority order)

Priority 1 — Must have before any drafting:

1. Source materials: Ask for the slide deck, brief, or docs to base the speech on. Read all provided files thoroughly.
2. Speech language: "What language will you deliver this speech in?" — The script will be written in this single language only. Do not produce bilingual versions unless explicitly requested.
3. Total duration: "How long is your speaking slot?" — Including or excluding Q&A.
4. Speaker persona: Name, title, organization, role in this event. "How should you introduce yourself?"
5. Core objective: "If the audience remembers one thing, what should it be?"

Priority 2 — Ask next, based on what's still missing:

6. Audience profile: Who's in the room — titles, seniority, organization, knowledge level.
7. Relationship context: First meeting or ongoing? What's been discussed before? What's the rapport level?
8. Event type: Keynote, roundtable, pitch, workshop, executive exchange?
9. Tone preference: Formal, conversational, data-driven, inspirational? See references/tone-and-style-guide.md for the full spectrum.
10. Speaking style reference (NEW): Present the style menu from references/speech-style-guide.md and let the user pick a style to reference. See details in the Speaking Style Reference section below.
11. Must-include / must-avoid: Topics that absolutely must or must not appear.

Priority 3 — Nice to have, ask if time allows:

12. Likely skepticisms: What doubts might the audience bring?
13. Case studies or data points: Specific numbers, stories, or examples to highlight.
14. Call to action: What should the audience do after the speech?
15. Other speakers: Who else is presenting? What are they covering?
16. Cultural considerations: Formality norms, humor appetite, directness level.

For the complete checklist with example prompts, see references/context-gathering-checklist.md.

Speaking Style Reference

This is a new feature that lets the user choose a famous speaker's style as a flavor reference for the speech script.

How to present: When asking Priority 2 questions, display the following category menu and ask the user to pick one (or skip):

🎤 演讲风格参考（可选）

你可以选择一位名人的演讲风格作为参考，我会在写台词时融入TA的节奏、用语习惯和叙事方式。

#	分类	可选风格
一	科技圈领袖	1. Elon Musk · 2. Steve Jobs · 3. Donald Trump · 4. Jeff Bezos
二	中国商业领袖	5. 马云 · 6. 任正非 · 7. 雷军 · 8. 周鸿祎 · 9. 刘强东
三	阿里系（按职级）	10. P7-P8 · 11. P9 · 12. P10 · 13. P11-P12
四	字节系	14. 张一鸣 · 15. 字节中高层通用
五	华为系	16. 任正非 · 17. 余承东 · 18. 华为中高层通用
六	腾讯系	19. 马化腾 · 20. 腾讯中高层通用
七	经典演说家	21. 马丁·路德·金 · 22. 奥巴马 · 23. 丘吉尔
八	新生代网红/意见领袖	24. 罗振宇 · 25. 俞敏洪 · 26. 李开复 · 27. 罗永浩

选一个编号，或回复"不需要"跳过。

After user selects: Read the corresponding style profile from references/speech-style-guide.md. Note the key attributes:

• Speaking tempo and rhythm
• Humor type
• Narrative structure
• Signature phrases and sentence patterns
• Emotional arc
• Vocabulary level

Store the selection as style_reference for use in Stage 3.

Briefing Guidelines

• Ask 3-5 questions per message, not all at once. Start with Priority 1.
• Accept answers in any format — shorthand, bullet points, info dumps.
• If the user provides reference documents, read them fully and extract: key themes, data points, organizational structure, implicit audience assumptions.
• Infer what you can from provided materials — don't ask questions already answered in the docs.
• When enough context is gathered (all Priority 1 items answered), confirm understanding with a brief summary and ask if anything is missing before moving to Stage 2.

Stage 2: Architecture

Design the speech structure before writing prose.

Select a Structure Pattern

Based on event type and objectives, select from the patterns in references/speech-structure-patterns.md:

Build the Timing Table

Create a section-by-section or slide-by-slide timing allocation:

| Section / Slide | Topic | Duration |
|----------------|-------|----------|
| S1 | Opening & Framing | ~3 min |
| S2 | [Topic] | ~2 min |
| ... | ... | ... |
| Total | | ~[X] min |

Use pacing guidelines from the structure patterns reference:

• ~130-150 words per minute spoken
• Chapter dividers: 0 min
• Data-heavy sections: 1.5-2x average time
• Buffer 15-20% for pauses and audience reactions

Define the Narrative Arc (Pyramid Principle)

Apply the McKinsey Pyramid Principle (Barbara Minto) to structure the narrative:

1. Top-down conclusion first: Open with the single overarching message — the "so what" — before presenting supporting evidence. The audience hears the answer, then the reasoning.
2. MECE grouping: Organize supporting arguments into 2-4 mutually exclusive, collectively exhaustive groups. Each group is a chapter/section of the speech.
3. Vertical logic: Each level answers "why?" or "how?" from the level above. If the top-level claim is "X is the right approach," each chapter proves one pillar of that claim.
4. Horizontal logic: Within each group, points follow either:
   • Deductive order (major premise → minor premise → conclusion)
   • Inductive order (similar facts → pattern → insight)
5. Situation–Complication–Resolution (SCR): Use this as the opening frame:
   • Situation: Common ground the audience already agrees with
   • Complication: The tension, change, or problem that disrupts the status quo
   • Resolution: Your key message / thesis (= the pyramid's apex)

Then fill in:

• Opening frame: SCR hook — situation the audience recognizes → complication that creates urgency → your resolution (the core message)
• Core thread: 2-4 MECE supporting pillars, each with its own evidence layer
• Closing frame: Restate the apex message, synthesize the pillars, and deliver the call to action

Present the architecture to the user for approval before proceeding to Stage 3.

Stage 3: Drafting

Write the complete speech script section by section.

Script Format

When input is PPTX — map each section to its corresponding slide number. This mapping is critical for Stage 4 (auto-injection into speaker notes).

For each section, produce:

## [Section Title] — Slide [N]

### Purpose
[One sentence: what this section accomplishes]

### Talking Points
- [Key point 1]
- [Key point 2]
- [Key point 3]

### Script
[Full spoken text, written in the designated language]

### Transition → Next Section
> "[Bridge sentence to the next topic]"

Writing Principles

1. One language only: Write the script in the language specified during briefing. Do not produce dual-language versions unless the user explicitly requests bilingual delivery.

2. Fit the persona: The script should sound like this specific person speaking — matching their seniority, style, and relationship with the audience. A CEO sounds different from a product manager. A first meeting sounds different from a fifth.

3. Apply the style reference (NEW): If a style_reference was selected in Stage 1, weave its characteristics into the script as a flavor layer:

   • Borrow 3-5 signature phrases or sentence patterns from the selected style — adapt them to the speech content, don't copy verbatim. For example, if the user chose "Elon Musk", transform a technical insight into: "What people don't realize is... [insight]. And that's... that's actually insane when you think about it."
   • Match the emotional arc: E.g., Musk = flat baseline → excited at vision moments; Jobs = controlled buildup → dramatic reveal; 马云 = storytelling warmup → inspirational climax.
   • Adopt the humor style: E.g., Musk's cold/meme humor, 周鸿祎's roast-style directness, 罗永浩's self-deprecating comedy, 雷军's wholesome data-nerd enthusiasm.
   • Mirror the vocabulary level: E.g., Trump = extremely simple repetitive words; 任正非 = literary/military metaphors; 张一鸣 = precise analytical language.
   • Use the narrative structure: E.g., Musk's "problem → current approach is absurd → first principles → solution"; Jobs's "buildup → One more thing → reveal"; 马云's "skepticism → reversal → deeper truth → golden quote".
   • Intensity calibration: If the user's actual seniority/context differs significantly from the chosen style (e.g., a junior engineer referencing "Alibaba P11-P12"), tone down the intensity by 50% while keeping the flavor. The style is inspiration, not cosplay.

4. Trust-building tone: Default to transparency and fact-based persuasion. Avoid over-promising. Acknowledge limitations honestly. Let data speak. See references/tone-and-style-guide.md for detailed guidance.

5. Self-Q&A technique: Where appropriate, pose questions the audience is likely thinking, then answer them. This creates engagement and demonstrates empathy. Use sparingly — 3-5 times per 30-minute speech.

6. Concrete over abstract: Use specific numbers, named examples, and tangible comparisons instead of vague claims.

7. Natural spoken rhythm: Short sentences (10-20 words average). Vary pace. Use bold for words to stress. Use em-dashes for pauses. Front-load key information.

8. Transitions are mandatory: Every section must end with a bridge to the next. Never have abrupt topic shifts.

9. Timing discipline: Each section's word count should match its allocated time at 130-150 words/minute.

Drafting Process

1. Draft sections sequentially following the approved architecture
2. After completing the full draft, do a coherence review:
   • Check for consistency in tone and terminology across sections
   • Verify timing adds up to the allocated total
   • Ensure transitions flow naturally
   • Confirm all must-include topics are covered and must-avoid topics are absent
   • Verify style reference consistency: If a style was selected, check that the chosen flavor is present throughout (not just in the opening) — signature phrases should appear in opening, middle, and closing sections
3. CRITICAL — Save to file: You MUST write the complete speech script to a .md file on disk (e.g., speech_script.md). Do NOT only output the content as chat text — the user needs a downloadable file. Use the write_to_file tool or equivalent file-writing mechanism to create the file.
4. Present the saved file to the user (show the file path and a brief summary)
5. Iterate based on feedback — use targeted edits, not full rewrites

Output Deliverables

The final output includes:

1. Speech script (markdown) — Complete section-by-section script with per-slide talking points and transitions
2. Timing overview table — Section-by-section time allocation
3. Speaker preparation notes (appendix) — Key data points to memorize, potential Q&A topics, tone reminders, and a summary of the applied style reference with 5 key phrases to practice

CRITICAL — File Output Rule: Every deliverable listed above MUST be saved as a file on disk (not just displayed as chat text). The speech script must be saved as a .md file. When a PPTX is provided, the injected PPTX and exported Word document must also be saved as files. Additionally, all files MUST be uploaded to a temporary file-sharing service (see Stage 4 Step 4 for details) to generate download links, because IM channels may not support file attachments from bots.

Stage 4: PPTX Injection (Auto-triggered)

This stage runs automatically when the user's input was a .pptx file. No need to ask — just do it.

What goes where

Design principle: Speaker notes in PowerPoint should be concise and scannable during presentation. Detailed preparation materials (timetable, Q&A, talking points, purpose) belong in the separate markdown/Word deliverable, not cluttering the notes pane.

Per-Slide Notes Format

For each slide, the injected speaker note should contain only the script and transition — clean, readable plain text:

[Script]
Full spoken text for this slide...

[Transition] → Bridge sentence to next section

Do NOT include [Purpose], [Talking Points], timing tables, Q&A preparation, or tone reminders in the PPT notes. Those belong in the markdown speech file.

Injection Workflow

1. Generate the notes JSON: After drafting is complete, produce a JSON file using a Python script (do NOT manually write JSON with write_to_file).

⚠️ Critical: When the speech content contains non-ASCII characters (Chinese, Japanese, special punctuation like "", ——), you MUST generate the JSON programmatically to avoid encoding issues:

import json
slide_notes = {
    "1": "第一页的演讲内容...",
    "2": "第二页的演讲内容...",
}
with open('notes.json', 'w', encoding='utf-8') as f:
    json.dump({"slide_notes": slide_notes}, f, ensure_ascii=False, indent=2)

This ensures all special characters are properly escaped.

The JSON structure:

{
    "slide_notes": {
        "1": "[Script]\nFull spoken text for slide 1...\n\n[Transition] → Bridge to next section",
        "2": "[Script]\nFull spoken text for slide 2...\n\n[Transition] → Bridge to next section",
        "3": "[Script]\nFull spoken text for slide 3..."
    }
}

Note: The JSON contains only slide_notes. There is no cover_note field. Each slide's value contains only [Script] and [Transition] — no Purpose, Talking Points, timing, or Q&A.

2. Run the injection script:

python scripts/inject_notes.py <original.pptx> <notes.json> [output.pptx]

The script path is relative to the skill root. Locate it with:

find . -name "inject_notes.py" -path "*/mck-speech-design-skill/*"

If the script is not available locally (e.g., skill loaded via GitHub URL), download it:

curl -o inject_notes.py https://raw.githubusercontent.com/likaku/mck-speech-design-skill/main/scripts/inject_notes.py

3. Verify (automated): The injection script includes built-in in-memory verification — it validates all notes are correctly set before saving, without re-reading the file from disk. If you need an independent post-hoc check:

from pptx import Presentation
prs = Presentation("output.pptx")
for i, slide in enumerate(prs.slides, 1):
    if slide.has_notes_slide:
        text = slide.notes_slide.notes_text_frame.text
        print(f"Slide {i}: {len(text)} chars ✅")
    else:
        print(f"Slide {i}: NO NOTES ❌")

If any slide shows ❌ or the file cannot be opened, fall back to manual delivery mode.

4. Deliver (CRITICAL — must not skip): After generating the output files, ensure all deliverables exist as files on disk, then upload each file to generate a download link so the user can retrieve them from any platform.

   File output checklist — verify each file exists before declaring success:

   • <name>_with_notes.pptx — the PPTX with injected speaker notes
   • <name>.docx — the Word document with full speech content
   • <name>_speech.md — the markdown speech script (generated in Stage 3)

   ⚠️ All deliverables must be saved as files, not just output as chat text. Always use file-writing tools (write_to_file, bash cat > file, Python file I/O, etc.) to save every deliverable to disk.

   CRITICAL — Generate download links: Many IM channels (QQ, 企业微信, etc.) do NOT support file attachments from bots. You MUST upload each deliverable to a temporary file-sharing service and provide the download URL in the chat. Use one of these methods (try in order):

   # Method 1: transfer.sh (preferred — 14-day retention, up to 10GB)
   curl --upload-file ./<filename> https://transfer.sh/<filename>
   # Returns a URL like: https://transfer.sh/abc123/filename.pptx
   
   # Method 2: file.io (single-download, auto-deletes after first download)
   curl -F "file=@./<filename>" https://file.io
   # Returns JSON with a "link" field
   
   # Method 3: 0x0.st (simple, no frills)
   curl -F "file=@./<filename>" https://0x0.st
   # Returns a direct download URL
   

   After uploading, present the results to the user in this format:

   ✅ 文件已生成并上传，点击链接即可下载：
   
   📊 演讲PPT（含备注）
   🔗 <download_url_for_pptx>
   
   📝 完整演讲稿（Word版）
   🔗 <download_url_for_docx>
   
   📄 演讲稿原稿（Markdown版）
   🔗 <download_url_for_md>
   
   ⏰ 链接有效期：14天
   

If injection fails

If the Python script encounters errors:

1. Fall back to manual mode: save the markdown speech script as a separate .md file
2. Save the notes JSON as a .json file
3. Upload all fallback files using the same curl --upload-file method described in Step 4 above, and provide download links to the user
4. Inform the user that notes could not be auto-injected, and present the download links for the fallback files

Export to Word (Automatic)

After PPTX injection, always also export the full speech markdown file to a .docx Word document. This is the complete reference document containing all preparation materials.

Run the export script:

python scripts/speech_to_docx.py <speech.md> [output.docx]

The script path is relative to the skill root. It converts the markdown speech file into a professionally formatted Word document with:

• Title and speaker info header
• Timing overview table
• Section-by-section script with clear headings
• Q&A preparation section
• Key data points appendix
• Speaker notes / tone reminders

The user receives two deliverables (saved as files on disk + uploaded with download links):

1. PPTX with speaker notes (Script + Transition only — scannable during presentation)
2. Word document with the full speech content (complete reference for preparation)

Delivery method: Since IM channels (QQ, 企业微信) may not support file attachments from bots, always provide download links (via transfer.sh or similar) in addition to saving files locally.