I Had 11 AI Subagents Test Every PPTX Skill for Claude Code. Here's What Works

Serge Bulaev ran eleven isolated Claude Code subagents against the same 5-slide investor pitch brief for Publora (a social publishing API spanning 10 platforms at $2.99 per account), each agent locked to one pptx skill’s own SKILL.md with no freelancing. Every deck went through the same LibreOffice-to-PNG render path and a human visual review. The run burned about 1.3M agent tokens and roughly one hour of wall-clock time; one skill’s instruction file alone (slides_maker) reached about 82,000 tokens before the agent touched the brief.

The decisive technical split was native OOXML versus painted layout. Slide 4 required a real table or chart, not an image of one. Only two pipelines produced native table/chart objects on their own: Anthropic’s official document skills (anthropics/skills) and slides_maker. Everything else, including the prettiest generators, built “tables” from text boxes and rectangles—looks fine in a screenshot, then breaks the moment someone tries to insert a row in PowerPoint. Agents scored each skill 1–10 on ease of use, output quality, editability, and docs; slides_maker’s strict lint refused to save a first build with 19 CRITICALs, and its render-linter reported contrast failures such as ink #E8EDF5 on fill #FFFFFF at 1.18:1. Anthropic’s skill carries a ~240-line rulebook of OOXML gotchas plus a render-and-look QA loop that even compiles a small C shim and LD_PRELOADs it so LibreOffice sockets work inside sandboxes.

For builders automating decks in agent pipelines, beauty and editability are currently a trade-off most skill authors resolved toward beauty. A deck that will be edited on a flight needs real tables, charts, and shapes you can mutate—not sixty-two loose objects where a table should be. Self-verification loops (build → lint → render → re-lint → fix) were the common trait of skills that survived contact with the brief; skills that trusted their first render failed hardest. That pattern matters beyond slides: agent tooling that cannot inspect its own artifacts will ship layout defects, white-on-white text, and format-incompatible content runs into production workflows.

The ecosystem already spans several product philosophies competing for the same Claude Code skill slot. Anthropic’s bundle treats .pptx as a document format with XSD validation and methodology that survived a missing pptxgenjs install by falling back to python-pptx. slides_maker ships a ~5,800-line helper with roughly 60 components and 14 design presets. dashi-ppt-skill (about 4,000 GitHub stars) locks ~1,020 layout pages and 12 themes for designer-grade covers but exports text boxes only and showed edge clipping under LibreOffice that its browser preview hid. mckinsey-pptx wins ease-of-use with about 40 templates and Harvey-ball consulting dialect, still as shapes. Image-first skills (ppt-image-first, gpt-image2-ppt-skills) score near the floor on editability by design and often depend on external image APIs. claude-office-skills turned out to be a verbatim pre-release snapshot of Anthropic’s skill—its README points upstream—so duplicate rows on a scoreboard can hide a single product line.

What to install in practice: anthropics/skills as the builder for native objects and QA; academic-pptx-skill as the editorial layer (ghost-deck action titles, exhibit annotations) paired with that builder; slides_maker when the input is a paper or codebase and you need native charts; mckinsey-pptx when the audience expects consulting dialect and edits go through code. Watch next for third-party skills that stop faking tables with rectangles and start emitting real OOXML—and for whether star-count leaders fix renderer fidelity and language-locked docs before teams adopt them as defaults.