Why we stopped letting AI design slides
Language models are good writers and unreliable designers. Here is how ViziSlides splits the work so every deck stays consistent.
The first wave of AI presentation tools asked one model to do everything: write the copy, choose a layout, pick colors, and position every box. The results looked impressive in a demo and fell apart in real use. Text overflowed. Slide five used a different font from slide six. A list of four items showed up as a pie chart.
We tried that approach and walked away from it. ViziSlides now splits the job in two, and the model only does the half it is good at.
Writers write, rules design
When you generate a deck, the model writes structured content for each slide: a headline, the points, the numbers, the quote. It never chooses a layout, a color, a font, or a position. That content then goes to a deterministic engine that reads its shape and picks the layout that fits.
- Three to six parallel points become cards or feature rows.
- Steps that happen in order become a process or timeline.
- Label and value pairs become stat cards or a KPI row.
- A series of values becomes a chart.
- Two sides of a choice become a comparison.
Because the same content always produces the same layout, your deck stays consistent. Because the theme is applied by code, not guessed per slide, every slide shares the same type, color, and spacing.
A model is a great first draft writer. It is a poor art director. So we stopped asking it to be one.
What you get from the split
- Predictable output. Regenerate a slide and the design stays put while the words improve.
- Real editability. Every slide is structured data, so you can switch layouts without retyping.
- Faithful exports. PowerPoint and PDF files are built from the same slide data you see in the editor.
- Fewer surprises. The model cannot invent a chart from numbers you never gave it.
