PPT Speech Script Generator

Generate a structured, oral-style speech script from an existing PowerPoint presentation.

When to Use

• User has an existing

  .pptx

  file and wants a speech script to go with it
• User says: "generate a speech script", "write speaker notes", "写演讲稿", "生成讲稿"
• This is NOT for creating PPT files — use the

  pptx

  skill for that

Workflow

Step 1: Extract Content

Extract text, visual layout, and embedded images from the presentation using three methods in parallel:

Text extraction:

python3 -m markitdown <pptx_file>

Visual layout (thumbnail grid):

python3 skills/pptx/scripts/thumbnail.py <pptx_file> <output_dir>/thumbnails --cols 5

Image extraction (unpack media files):

python3 skills/pptx/ooxml/scripts/unpack.py <pptx_file> <output_dir>/unpacked

Then:

1. Read the thumbnail grid image(s) to understand the visual structure of each slide
2. Build a slide-to-image mapping by reading the relationship files:
   • Parse

     ppt/slides/_rels/slideN.xml.rels

     for each slide
   • Extract

     Target="../media/imageX.ext"

     references to find which images belong to which slide
3. For slides with important images (charts, architecture diagrams, screenshots), read the original image from

   ppt/media/

   using the Read tool for detailed visual analysis
   • Skip generic/decorative images (backgrounds, logos, icons) — focus on content-carrying images
   • Prioritize: architecture diagrams, data charts, comparison tables, workflow diagrams, product screenshots

Image analysis guidelines:

• For architecture/flow diagrams: describe the components, data flow direction, and key relationships
• For data charts: read exact numbers, axis labels, and trends
• For comparison tables: extract the key differentiators
• For product screenshots: describe what the user interface shows
• For photos/decorative images: brief description only, don't over-analyze

This image analysis enables the speech script to accurately describe visual content that

markitdown

cannot capture (since markitdown only extracts text, not image content).

Step 2: Analyze Presentation Structure

Before writing, analyze the deck:

1. Identify sections — group slides into logical sections (intro, body sections, conclusion)
2. Identify hidden slides — thumbnail.py reports hidden slides; exclude them from the script
3. Identify slide types — title, content, data, diagram, comparison, section divider, closing
4. Note visual elements — charts, images, diagrams that need verbal explanation
5. Estimate timing — allocate time per slide based on content density

Timing guidelines:

Slide type	Suggested time
Title/cover	30s - 1min
Agenda	30s - 1min
Section divider	15 - 30s
Content (light)	1 - 1.5min
Content (dense)	1.5 - 2.5min
Data/chart	1.5 - 2min
Diagram/architecture	2 - 3min
Demo/code	2 - 3min
Summary	1 - 2min
Closing/Q&A	30s

Step 3: Write the Speech Script

Generate a complete speech script following these principles:

Output Format

---
title: "PPT Title — Speech Script"
slides: <total visible slides>
estimated_time: "XX-XX minutes"
audience: "<target audience>"
---

# PPT Title — Speech Script

> Audience: <target audience> | Date: <date if available>

Estimated duration: XX minutes (including Q&A). Suggested time per slide is noted in brackets.

---

## Slide 0 — Slide Title [30s]

Speech content here...

---

## Slide 1 — Slide Title [1min]

Speech content here...

---

...

## Predicted Q&A

### Q1: <likely question>?
Answer key points...

### Q2: <likely question>?
Answer key points...

### Q3: <likely question>?
Answer key points...

Writing Principles

1. Oral style — Write as if speaking to the audience, not reading a document

   • Use conversational connectors: "Let's look at...", "The key takeaway here is...", "Now, moving on to..."
   • Avoid academic or written-style phrasing

2. Don't read the bullets — The script should EXPLAIN and EXPAND on slide content, not repeat it

   • Slide says "Cost reduced 34.8%" → Script says "We brought per-query cost down by over a third — from 9 cents to under 6 cents. At million-user scale, that's millions of dollars saved per month."

3. Per-slide length: 80-200 words (Chinese) or 60-150 words (English)

   • Section dividers and title pages: shorter (30-60 words)
   • Dense data or architecture slides: longer (150-250 words)

4. Smooth transitions — Each slide's script should naturally flow from the previous one

   • End of previous slide's conclusion → Beginning of next slide's topic
   • Use transitional phrases: "That brings us to...", "With that context in mind...", "So how do we solve this?"

5. Highlight key points — Use bold or verbal cues for emphasis

   • "The most important number on this slide is..."
   • "If you remember one thing from today..."

6. Explain visuals — For charts, diagrams, and images, guide the audience through what they're seeing

   • "Looking at this architecture diagram, data flows from left to right..."
   • "The blue bars represent the baseline, and the orange bars are our optimized results..."

7. Audience awareness — Tailor depth and terminology to the stated audience

   • CTO audience → focus on strategic impact, cost, and scalability
   • Developer audience → focus on implementation details and code
   • Business audience → focus on ROI, user impact, and market context

8. Q&A section — Prepare 3-5 predicted questions

   • Include "tough but fair" questions the audience is likely to ask
   • Provide concise answer key points (not full scripts)
   • Consider the audience's perspective and concerns

Language

• Default: Match the language of the PPT content
• User override: If user specifies a language (e.g., "in Chinese", "in English"), use that
• Mixed content: If PPT has mixed languages, use the dominant language unless told otherwise

Step 4: Save Output

Save the speech script to the same directory as the PPT file:

# Output path: same directory as input, named speech-script.md
<pptx_dir>/speech-script.md

If a

speech-script.md

already exists, ask the user before overwriting.

Dependencies

These should already be available from the

pptx

skill:

• markitdown:

  pip install "markitdown[pptx]"

  — text extraction
• thumbnail.py:

  skills/pptx/scripts/thumbnail.py

  — visual layout analysis
• unpack.py:

  skills/pptx/ooxml/scripts/unpack.py

  — PPTX unpacking for media extraction
• LibreOffice: for PDF conversion (used by thumbnail.py)
• Poppler: for PDF-to-image conversion (used by thumbnail.py)

Edge Cases

Very large presentations (50+ slides):

• Group consecutive similar slides into sections
• Summarize repetitive slides rather than scripting each individually
• Note in the script: "Slides X-Y cover [topic] — walk through highlights"

Image-heavy / text-light slides:

• Use the slide-to-image mapping to read original images from

  ppt/media/

  at full resolution
• Describe what the audience sees based on direct image analysis
• For images that cannot be read (e.g., unsupported format), fall back to thumbnail grid analysis
• Flag slides where content is still unclear: "[Note: This slide contains a visual element — verify description against actual slide]"

Hidden slides:

• Exclude from the main script
• Optionally note them at the end: "Note: Slides X, Y, Z are hidden and not included in this script"

No text content (pure image deck):

• Extract all images via unpack and read each one directly for full-resolution analysis
• Use thumbnail grid for overall slide layout understanding
• Generate descriptive narration based on per-image analysis
• Clearly note which descriptions are based on visual interpretation

Quality Checklist

After generating the script, verify:

• Every visible slide has a corresponding section
• Slide numbering matches the PPT (0-indexed)
• Transitions between slides are smooth
• No bullet points are simply repeated verbatim
• Time estimates per slide are reasonable
• Total estimated time is realistic for the slide count
• Q&A section includes 3-5 relevant questions
• Language matches user's request or PPT's dominant language