Choosing an AI PPT skill: HTML or native pptx

One premise first: in 2026, searching for "PPT skills" almost always lands you in the Claude Code / Codex agent ecosystem — skill packages in the SKILL.md format — not traditional advice on making better-looking slides. That space now has its own leaderboards, review sites, and a lot of discussion.

After digging through it, the honest conclusion is this: there is no ranking that every source agrees on.

The winner by install count, the winner by GitHub stars, and the winner by independent scoring are three different things. More importantly, these skills do not produce the same artifact. Some emit a real .pptx you can edit in PowerPoint; others emit an HTML web deck. Those two do not compete on a single axis — they solve different problems.

The official baseline: Anthropic's own pptx skill#

Claude on the web, Claude Code, and the API all ship a set of official skills covering docx, pdf, pptx, and xlsx. The pptx one can build a full deck from an outline or topic, follow corporate templates and master styles (brand colors, layouts), insert text, images, tables, and charts, add or remove speaker notes, and extract text from an existing pptx for summarization or analysis. It runs on python-pptx underneath.

The anthropics/skills repo it lives in has roughly 172k stars, but the licensing splits in two. Example skills like skill-creator and mcp-builder are fully open under Apache 2.0. The production skills that actually drive Claude's document capability — docx, pdf, pptx, xlsx — are labeled "source available, not open source," published as a reference for developers. The pptx folder's license reads Proprietary.

More significant: the Skill standard itself has been adopted as a common format by roughly 40 mainstream agent tools, including GitHub Copilot, VS Code, Cursor, OpenAI Codex, and Gemini CLI. So if "consensus" means "the most widely adopted standard implementation," the answer is this one — it just isn't something you install.

By install count: every leader has a catch#

One third-party leaderboard ranks by actual install count. It indexes over 42,000 skills, 179 of them under a "presentations" category. Data as of 6 July 2026:

The problem with this list is blunt: the install-count leader, lark-slides, is a Feishu-only tool. Outside that ecosystem it is useless to you. Install count here measures ecosystem size, not quality.

The two with the strongest reputation in the Chinese community#

guizang-ppt-skill and huashu-design drew the most independent discussion I could find, and they are frequently benchmarked against each other — Bilibili has hands-on comparison videos of exactly this pair.

guizang-ppt-skill (@op7418) currently sits at 22.1k stars and 1.6k forks under AGPL-3.0. It converts an article into a single-file HTML horizontal deck, with two built-in visual systems — "digital magazine" and "Swiss internationalism" — and 22 locked layouts. The design philosophy is deliberately restrictive: theme colors cannot be customized, on the stated grounds that "protecting the aesthetic matters more than granting freedom." WebGL backgrounds appear only on the first screen. Images must align to the body text column.

The author draws the boundary himself: good for offline talks, internal presentations, product launches, demo days — anywhere a strong personal voice helps. Bad for dense tabular data, training material (not enough information density), and anything requiring collaborative editing. Its deliverable is HTML, not pptx.

huashu-design (@AlchainHust) is a broader HTML design skill — it also does prototypes and animation — but decks are one of its primary uses. GitHub snapshots I found put it between 22k and 23k stars; it is growing fast enough that different capture times disagree.

Its technical hook is html2pptx.js, which reads each element's computed style from the DOM and converts it into a PowerPoint object, so the export contains real text boxes rather than pasted screenshots. If that holds up, it is one of the few skills that gets a genuinely editable pptx out of an HTML pipeline.

The author is candid about the limits, calling it "an 80-point skill, not a 100-point product": no layer-level editable PPTX export to Figma, nothing approaching Framer Motion-grade 3D or physics animation, and quality drops to 60–65 with no brand input at all.

Licensing is contradictory across sources. One hands-on Zhihu piece says MIT and commercially usable; a later independent analysis says commercial use still requires explicit permission from the author. Check the repo's current LICENSE yourself before shipping anything with it.

One trap worth naming#

A search will surface "10 Best AI PPT Skills for Claude Code CLI (Ranked & Tested)," complete with a rigorous-looking six-dimension scoring table. The winner is Felo Slides. The article is published by Felo — the winner is their own product, and the official Claude pptx skill scores poorly in it. Classic self-promotion dressed as a review. Borrow the framework if it's useful; ignore the ranking.

The other kind of noise is name collision. frontend-slides maps to entirely different repos depending on the source: one lists 25k stars under MIT, another leaderboard attributes it to a different repo showing over 230k stars. Disagreements like this are common here. Open the actual GitHub repo and verify the author and star count yourself rather than trusting an aggregator.

If the hard requirement is "editable in PowerPoint"#

State that constraint and most of the list above drops out:

• guizang-ppt-skill's own FAQ concedes that "the current core deliverable is HTML… if you need PPTX, treat the HTML as a visual comp and convert it, but that is not the current main path"
• huashu-design, frontend-slides, Slidev, and Marp all produce HTML or PDF as their primary artifact
• lark-slides and gws-slides target Feishu and Google Slides formats, not Microsoft pptx

For the candidates that do clear the bar, I cross-checked against a separate third-party aggregator that scores skills on security and quality. It tracks 30 PPT-related skills totaling 133.3k stars and refreshes every 8 hours:

The quality score is that aggregator's composite across six dimensions (code completeness, documentation, maintenance activity, and others), not my own assessment. Treat it as a signal, not a verdict.

Two warnings about ppt-master#

ppt-master positions itself as producing "a real, editable PowerPoint from any document — native shapes and animations, speaker notes voiced as audio narration, and the option to follow your own .pptx template, not slide images." Three independent sources list it as the benchmark for real pptx output in this category, and it holds the top composite score among all 30 PPT skills on that scored leaderboard.

Two things need saying anyway.

First, "python-pptx can read it" is not "PowerPoint can open it." A real GitHub Discussion (#40, June 2026) records a user following the Windows install guide to generate a 3-slide test deck. The export reported success and python-pptx read all three slides fine — but Microsoft PowerPoint itself reported the file as corrupt. The cause: the notes page's relationship file referenced a notesMaster.xml that did not exist in the package. The author replied the same day, confirmed the report was useful, and shipped a fix by removing master notes.

That is exactly where the "native editable pptx" requirement actually breaks. A script validating its own output and Microsoft's parser tolerating that output are separate questions. Open the result in real PowerPoint — or at minimum LibreOffice — rather than trusting the tool's own "generated successfully."

Second, GitHub hosts a completely different project under the same name: woqianfu/ppt-master, a decision framework integrating 22 different PPT skill paths, which itself lists hugohe3's ppt-master and MiniMax's PptxGenJS approach as options. Don't grab the wrong repo.

What I'd pick, by scenario#

Client deliverables or embedding into your own product: I'd try Anthropic's official pptx skill first. It ships inside Claude with nothing to install, and precisely because it doesn't chase visual flourishes, it spends its effort on not producing corrupt files — generation is followed by mandatory schema, relationship, and chart validation scripts, plus a requirement to render each slide to an image for human visual review. That generate → validate → visual QA loop is stricter than most community skills.

ppt-master's edge is the things the official skill doesn't emphasize: native animation, transitions, voiced narration, and applying your existing pptx template. The two don't conflict. Produce a clean base with the official skill, then hand it to ppt-master when you need the motion layer.

Everything else:

• Visual impact matters, format doesn't: guizang-ppt-skill (Swiss / magazine, restrained) or huashu-design (broader HTML design range)
• Your company already runs on Feishu: lark-slides
• Conference talks for a technical audience, with embedded code and live data: Slidev (antfu's team) or Marp. Be clear that Slidev outputs an interactive web page and cannot export to PowerPoint format
• Papers, lab meetings, thesis defenses: nature-paper2ppt is a niche skill built for exactly this

One closing note: star counts and license terms in this ecosystem change daily. Every number above is a snapshot from when it was captured. Verify before you install.