What's this about?
dotey's argument: having AI *operate* PowerPoint is the wrong shape — not impossible, but the results are ugly, because current models are good at generating, not at operating. Inventing an intermediate YAML format, as Kimi did with PPTD, is workable but insufficient for two reasons: the format has no training behind it, so Kimi needs nearly 2,000 lines of documentation to explain it, and YAML's expressiveness is too limited to describe full formatting. Templates fix appearance partially while capping the model's imagination. He invokes the bitter lesson — the brute-force method that uses compute and general learning beats the clever shortcut built on human expertise — and concludes the best current practice is Claude Design's: define the style as a design system, generate HTML, then export to editable PPTX.
Why is HTML the right intermediate format?
Because the model has seen more of it than of anything else with visual layout. Design capability in these models is real but drops sharply for non-standard formats, and HTML is the one visual language they are fluent in. PPTX is a standard, so the export is a solved mechanical problem — dotey points to pptxgenjs, and has built his own version called gen-pptx. The chain works because each step plays to something's strength: the model generates in a language it knows, and a library handles the format conversion.
What's the community saying?
- The strongest complementary argument comes from YinsenW_: the problem is not the model but the tooling — PowerPoint is closed-source, complex, and exposes no complete set of atomic operations suitable for an AI to call, so everything is built on open-source Office libraries or hand-assembled toolchains that don't match PowerPoint even for humans.
- His conclusion follows: if you want AI-made slides to work, build a genuinely AI-native slide tool rather than optimising the model further.
- A practical question in the replies that nobody fully answers: how hard is the HTML-to-editable-PPTX step really?