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

I'll pull the article and HN thread so the paragraphs stay grounded in real details rather than invented figures.Serge Bulaev ran eleven isolated Claude Code subagents against the same five-slide investor pitch brief for Publora, a social publishing API priced at $2.99 per account. Each agent cloned one PPTX skill repo, followed that skill’s SKILL.md without freelancing, and scored ease of use, output quality, editability, and docs on a 1–10 scale. The run burned about 1.3M agent tokens in roughly one hour of wall-clock time, produced eleven decks, and rendered every file through one LibreOffice-to-PNG pipeline before a human visual review. On Hacker News the write-up sits at 1 point with 0 comments.

The hard trap in the brief was slide 4: a real OOXML table or chart, not a picture of one. Only two of eleven skills produced native table or chart objects through their own pipeline — Anthropic’s official document skills and slides_maker. Almost every third-party skill drew “tables” from text boxes and rectangles; one pretty comparison matrix was sixty-two loose shapes. slides_maker’s stack is a roughly 5,800-line python-pptx helper with about 60 components, 14 design presets, and a build → strict-lint → render → render-lint → critic loop that refused to save the first build with 19 CRITICALs and later caught 30-plus defects. Its SKILL.md alone can run to 82,000 tokens. Anthropic’s skill is shorter (~240 lines of gotchas) but treats .pptx as a document format: real tables and charts, an XSD validator, and a render-and-look QA path that can compile a C shim with GCC and LD_PRELOAD it so LibreOffice sockets work in sandboxes. dashi-ppt-skill (about 4,000 GitHub stars) looked designer-grade with ~1,020 pre-built layout pages and three validators, yet exported roughly forty absolutely positioned text boxes for the comparison “table” and clipped text under LibreOffice even when the browser preview looked fine.

For engineers shipping agent-built decks into real workflows, beauty versus editability is the live trade-off. A screenshot-perfect deck that opens as loose shapes fails the moment someone adds a row on the plane. Skills that survived contact with the brief shared a self-verification loop: render, inspect, fix, repeat. Skills that trusted first-pass output failed hardest — ppt-agent-skill needed fifty-five minutes of fixes for a valid deck, including SVG converters that dropped gradients into white-on-white and “18 charts” that were HTML/CSS recipes, not chart objects. If the file will be edited in PowerPoint, native OOXML objects and a render-QA loop matter more than star counts or consulting aesthetics.

The field has gone from a curiosity to an ecosystem in about six months: Anthropic’s official bundle, McKinsey-style template packs with Harvey balls, academic editorial methodologies, thesis-defense pipelines bound to Windows PowerPoint COM, and image-first tools that wrap PNGs in a PPTX shell and score a 2 on editability. claude-office-skills is a verbatim pre-release snapshot of Anthropic’s skill and should be treated as the same entry. mckinsey-pptx won highest ease of use (plan → pick from 40 templates → build in about one second of Python) but still ships shape-drawn matrices. Image-model skills such as ppt-image-first and gpt-image2-ppt-skills are narrow by design: no external image API means no output.

What Bulaev would actually install is a small stack: anthropics/skills as the builder for native objects, validation, and QA; academic-pptx-skill as the editorial layer (ghost-deck action-title tests and exhibit annotations); slides_maker when the input is a paper or codebase and charts matter; mckinsey-pptx when the audience expects the consulting dialect. Watch whether third-party skills stop painting tables and start emitting real OOXML, whether dashi-style beauty engines close the LibreOffice fidelity gap, and whether skill authors keep shipping mandatory render-inspect loops instead of trusting first-pass output.