wtf is deck-dna?

deck-dna is a Claude skill that helps an AI assistant build PowerPoint slide decks that actually look like they came from a real designer, instead of the generic "AI template" look that gives machine made decks away even when the colors and logo are technically correct. The approach works by measuring, rather than describing, the visual style of decks you already trust. You give it two or three approved decks that a real designer built and a client signed off on, and the skill extracts what it calls a design DNA sheet: where images sit compared to text, which colors are actually used on the slides rather than what the file's theme claims, how a large statistic is styled compared to a small one, and when divider slides tend to appear. A small script also counts words per slide across those reference decks, and that measured number becomes a hard limit for new decks, so the fix for too much text is cutting content rather than shrinking the font. After building a new deck, the skill converts it to PDF, renders each slide as an image, and has the agent look at its own output to catch problems like broken text, overlapping elements, or crooked margins, then fix and re-render as needed. To use it with Claude Code, you download a single SKILL.md file into a skills folder, then ask Claude to build a deck using your reference decks as the style guide. It requires Python along with the python-pptx and Pillow libraries to build the deck, and LibreOffice plus Poppler if you want the automatic visual review step, though it can fall back to manual review without those. The README is honest about its limits: it only counts words in text boxes and not tables, font embedding can be finicky, and rendered checks using LibreOffice are only approximate since it may substitute fonts. The project is released under the MIT license.