I have been building MDPR, a deterministic Markdown-to-editable-PPTX runtime, and mdpr-skill, an optional Codex companion for review hints.
The problem I wanted to avoid is common in LLM-generated slide decks: the model owns too much. It chooses coordinates, colors, emphasis, spacing, and object geometry directly. The first slide may look plausible, but coherence often drifts across the deck.
MDPR uses a different boundary:
- Markdown is parsed into presentation structure.
- Slide splitting and layout planning are deterministic.