// the find
hugohe3/ppt-master
AI generates a real, editable PowerPoint from any document — native shapes & animations, editable charts & tables you can change the data on, speaker notes voiced as audio narration, and the option to follow your own .pptx template, not slide images · by Hugo He
PPT Master is a Python workflow that runs inside AI coding environments (Claude Code, Cursor, etc.) to generate natively editable .pptx files from documents, PDFs, or pasted text. The core technical bet is using SVG as an intermediate representation — the AI generates slide layouts as SVG, which is then converted to real DrawingML shapes rather than flat images. It is aimed at anyone who needs presentation decks and has already bought into an AI IDE.
The SVG → DrawingML pipeline is the right call. AI models have strong SVG knowledge, and converting to real shapes instead of embedding per-slide screenshots is what makes the output actually useful — the examples show genuine slide structure, not image wrappers. The --native-objects flag emits real PowerPoint chart and table objects backed by editable data, which almost no AI presentation tool bothers to implement. Model and platform agnostic by design: works with Claude, GPT, Gemini, and any IDE with file system access. The audio narration pipeline is a full production feature — speaker notes to TTS to embedded audio in the PPTX — not a stub.
This is not software you install and call; it's a workflow that requires a running AI IDE in the loop for every generation. There's no CLI, no API, no way to integrate it into a CI pipeline or call it from application code. Quality is entirely model-dependent, and the recommended setup (Claude Opus plus gpt-image-2) is expensive — a complex deck with many images can easily cost $10–$20+ per run, and the README's honest warning about cheaper models producing worse results means the cost ceiling is real. The --native-objects chart output carries a caveat that rendering may vary across PowerPoint, Keynote, and WPS — so you're trading visual fidelity for data editability in a way that may bite you at presentation time. The ~1M token context recommendation effectively constrains you to Claude or Gemini for non-trivial documents.