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

I'll pull the source article so the paragraphs stay grounded in real claims rather than inventing details from the title alone.Serge Bulaev published a head-to-head of eleven Claude Code presentation skills, each driven by its own isolated subagent on one fixed brief: a five-slide investor pitch for Publora, a real social publishing API priced at $2.99 per account across ten platforms. Slide 4 had to include a real table or chart, not a picture of one. Agents cloned a skill repo, followed only that skill’s SKILL.md, scored ease of use, output quality, editability, and docs on a 1–10 scale, and wrote reports on failures and workarounds. All decks went through the same LibreOffice-to-PNG path for visual review. The run cost roughly 1.3 million agent tokens and about an hour of wall-clock time.

The hard result cut across the field: almost every third-party skill painted tables instead of building them. Only two of eleven produced native OOXML table or chart objects through their own pipeline—Anthropic’s official document skills and slides_maker. The rest, including several of the prettiest decks, faked tables with text boxes and rectangles; one McKinsey-style comparison was shapes that looked fine in a screenshot but would not take a new row in PowerPoint. Beauty and editability traded off, and most authors chose beauty. Instruction files themselves are now large systems: slides_maker’s SKILL.md ran to about 82,000 tokens; the official skill’s shorter rulebook still carried a fallback from missing pptxgenjs to python-pptx without collapsing. Strong skills shared a build–lint–render–inspect loop; slides_maker’s strict linter refused a first save with 19 criticals, and its render linter reported contrast failures with concrete hex and ratio numbers.

For engineers shipping agent-built decks, the scoreboard is a workflow filter, not a beauty contest. If someone will edit the file in PowerPoint, native tables, charts, XSD validation, and a render-QA loop matter more than Harvey balls drawn from masked circles. dashi-ppt-skill (thousands of GitHub stars) can look designer-grade but exports text boxes only and clipped text under independent LibreOffice rendering; mckinsey-pptx scored highest on ease of use with plan-and-template flow, yet still zero native table objects. Image-first skills that wrap PNGs in a PPTX shell scored near the bottom on editability by design and often hard-depend on paid image APIs.

The market for Claude Code pptx skills is already fragmented: Anthropic’s upstream document skills, a GitHub snapshot that is effectively the same skill, consulting cosplay packs, academic editorial methodologies, thesis-defense pipelines tied to Windows PowerPoint COM, and Chinese-first or image-model-only tools. Some repos market broadly while shipping narrow stacks—missing layout references, hardcoded paths, language-locked SKILL.md files, or SVG converters that drop gradients and polygons. Star counts and screenshot polish do not predict whether the file is a document or a picture of a document.

Install for the job, not the leaderboard. Bulaev’s practical shortlist is anthropics/skills as the builder for native objects and QA, academic-pptx-skill as the editorial layer (action titles, ghost-deck outline test, exhibit annotations) when paired with that builder, slides_maker when the input is papers or code and you need real charts, and mckinsey-pptx when the audience expects consulting dialect. Watch whether skill authors ship real OOXML tables and charts, portable dependencies, and mandatory self-verification loops; skills that trust first output without a render-and-fix pass failed this brief hardest. Full per-skill reports and all fifty-five slide renders are available from the author on request.