I make getdeck, a slide layout library (https://getdeck.co). The slides are HTML, and people need them in PowerPoint, so I built an export to .pptx. I also designed a self-contained .html export, which is not on the site yet. A few things I learned along the way:
1. Measure the slide, then rebuild it. The export renders each slide at a fixed 1920x1080, records where every element actually landed, and builds the .pptx from those coordinates. The UI that asks for an export never has to know about positions or styles.
2. Not everything can stay native. Text stays editable and charts become real PowerPoint shapes, but some decorations only survive if they are baked into an image. When I design a new layout now, "this part will become a picture" is a warning sign.
3. A round trip needs two fingerprints. In the HTML export design (not live yet), the file carries its coordinate data inside, so it could be edited and uploaded back. With a hash of that data alone, editing only the visible text would still match and the old text would go out. So the design checks the data and the visible content separately, and measures the slide again when either one changed.
AI writes decent slide copy and terrible decks. getdeck hands your model the missing half: 154 layouts and 41 themes it can reason about over MCP, plus the rules for joining them into one coherent deck.
No repeats, no drift.
