I Tested 11 Claude Code PPTX Skills With AI Subagents — Results

An engineer ran a controlled experiment testing eleven third-party PPTX-generation skills for Claude Code, using isolated AI subagents that each followed their skill's own documented methodology on an identical investor-pitch brief. Every resulting deck was rendered through the same LibreOffice-to-PNG pipeline for visual, apples-to-apples comparison — consuming roughly 1.3 million agent tokens over about an hour.

The central finding: only two of eleven skills — Anthropic's official document skill and slides_maker — produced genuinely native, editable OOXML tables and charts. The rest, including the most visually polished generators, fake tables using loose text boxes and shapes that look identical on screen but fall apart the moment someone tries to edit a row in PowerPoint. This exposes a real trade-off in the ecosystem between visual polish and structural editability.

Other entries tested included a visually striking tool whose fidelity depends heavily on its own proprietary renderer, a template-locking system with Chinese-only documentation, and a Korean-documented tool that nails consulting-deck aesthetics but sacrifices native table support. For engineers building automated deck-generation pipelines, the takeaway is concrete: pick the official skill or slides_maker for anything that will be edited afterward, and reserve the flashier third-party renderers for decks meant purely for presentation.