PPT Master Skill Deep Dive: SVG Intermediary, Web Annotations, and Native Editable PPTX

Most AI slide generation tools simply paste full-bleed flat images onto slides. While visually appealing initially, text cannot be edited, shapes cannot be moved, and zooming in results in blurry rasterization.

ppt-master is an open-source Skill designed for AI Coding Agents (such as Claude Code, Qwen Code, and Codex). Its core value lies in parsing your documents to produce fully native DrawingML PPTX files where every text box can be edited and every vector shape can be recolored and adjusted.

1. End-to-End User Experience

The typical collaboration workflow with this Skill is seamless:

1. Triggering the Skill:
   • Explicit Command: Trigger directly via slash commands (e.g. /ppt-master <doc-path>) or variable injection ($ppt-master);
   • Natural Language Intent: Say naturally: “Convert this technical spec into a 10-slide presentation.”
2. Automated Structuring & Vector Layout: The AI agent extracts document hierarchy, designs visual outlines, and generates SVG vector code for each slide.
3. Local Web Preview, In-Place Annotations & Precise Refinement:
   • A local Flask server runs at http://127.0.0.1:5050 rendering the generated SVGs;
   • You can review slides in the browser, click on any unsatisfying text or shape to add an annotation, and hit Submit Annotations to write changes to disk;
   • Return to your terminal and instruct the AI: “Annotations submitted, please refine.” The AI inspects disk annotations via check_annotations.py and regenerates only targeted elements.
4. Export Native PPTX: Once satisfied, the final SVGs are compiled locally into a native .pptx file ready for secondary editing in Microsoft Office or WPS.

Ecosystem Comparison

2. Architecture: Why Are Slides Truly Editable?

The Two-Stage Conversion Architecture (SVG → DrawingML)

PowerPoint relies on Microsoft’s DrawingML XML specification. Direct generation of DrawingML by LLMs frequently suffers from XML syntax errors and broken files.

ppt-master resolves this through a two-stage intermediary model:

Source Document → [AI Vector Layout] → SVG Code → [Deterministic Python Script] → DrawingML (PPTX)

• Stage 1 (AI Creative Layout): LLMs naturally excel at producing clean SVG markup, handling visual layout, palettes, and typography;
• Stage 2 (Deterministic Compilation): A local Python script (svg_to_pptx.py) maps SVG tags (<text>, <rect>, <path>) directly into native PowerPoint DrawingML XML nodes.

3. Security Audit & Zero-API Fallback

1. Security Audit Findings

• Network Isolation: Local preview binds strictly to 127.0.0.1:5050 with zero telemetry;
• Execution Safety: Zero eval() or exec() usage; subprocess calls only internal deterministic toolchains;
• File System: Reads and writes are scoped to project paths with path-traversal safeguards;
• Verdict: Clean and safe.

2. Zero External API Dependency & Offline Manual Mode

The core generation pipeline runs completely offline without commercial APIs:

Offline Manual Fallback: If no image generation API key is configured, image slots are automatically tagged as Needs-Manual with dashed prompt placeholders, ensuring the PPTX export pipeline finishes without blockers.

4. Installation & Dependency Decoupling

1. Handling npx skills add Timeouts:
   • The repository contains ~99MB of template assets. If the CLI times out, use rsync to backfill scripts/ (105 Python scripts) from a full clone.
2. Bypassing PEP 668 via uv venv:
   • Modern Linux distros block global pip installation. Use uv to establish a clean, isolated virtual environment:

     uv venv ~/.ppt-master/venv
     uv pip install -r requirements.txt --python ~/.ppt-master/venv/bin/python
     

3. Decoupling pycairo / svglib:
   • svglib (requiring system libcairo2-dev) is merely an optional fallback for legacy Office rasterization. Core DrawingML export does not require it and runs smoothly with the remaining 88 dependencies.