Humanize PPT: an agent skill that directs the outline before the slides exist

The outline is the product

The repository opens with a sentence that doubles as its thesis: template libraries can spread one concept across a dozen pretty HTML pages, but what you have to deliver, standing, is the line that pushes the audience forward page by page. Humanize PPT is built on the far side of that observation.

Its method is named AST, audience state transitions: material is planned into a line where every page advances what the audience understands by one increment. The author's framing comes from experience, having given many talks and finding that the attractive HTML deck skills suit concept display, where a single simple idea gets padded across a dozen pages, while a ninety-minute talk supports only thirty-some pages and needs every one of them to carry.

The division of labour follows: this skill does the work template libraries do not, deciding the line, the intent of each page, and where real visuals belong, while downstream skills handle rendering. It is a director layer, and the repository's structure, a skill with contracts and adapters, reflects that role.

Install, and the downstream constellation

Installation follows the Agent Skills pattern, with the simplest path being handing your agent the repository link:

请安装 Humanize PPT Skill：https://github.com/LearnPrompt/humanize-ppt

Or the one-line package route:

npx skills add LearnPrompt/humanize-ppt -g

Claude Code users get a marketplace path with automatic updates:

/plugin marketplace add LearnPrompt/humanize-ppt
/plugin install humanize-ppt

The skill deliberately does not render slides itself. A table of downstream skills covers that: a Chinese-deck native renderer in magazine and Swiss styles, an English viewport-safe HTML renderer, an English multi-template renderer, a native PPTX producer working in DrawingML with notes, transitions and raw template fill, a Remotion-based orchestrator producing real MP4 explanation videos, and an image-generation skill that runs through a local Codex CLI without an API key. Pairing advice is given from practice: use the Remotion orchestrator for flow, a best-practices skill to guard unstable patterns, and a larger toolkit only for complex engineering like subtitles, charts and batch pipelines.

That constellation is the honest architecture: an outline director is only as good as the renderers beneath it, and the README treats choosing them as part of the setup rather than an embarrassment.

The one-sentence workflow

The documented usage is a single prompt, quoted verbatim in Chinese:

用 humanize-ppt 把这份材料做成中文演讲 PPT：先出 AST 大纲和每页意图，
按大纲调 guizang-ppt-skill 原生渲染，配图用 baoyu-image-gen、视频用 remotion，
渲染完跑一遍演讲体检告诉我哪几页只能看不能讲，最后出演讲模式。

Translated: use humanize-ppt to turn this document into a Chinese presentation deck, producing the AST outline and each page's intent first, then native rendering through the Chinese renderer per the outline, images through the local image skill, videos through Remotion, and after rendering run the presentation checkup telling me which pages can only be looked at and not spoken, finally producing presentation mode. The English variant swaps the language and renderer; beginners are told to ignore the staged CLI controls below that prompt.

For native PowerPoint output, the sentence changes shape: humanize-ppt defines the AST, the PPTX producer renders natively, and the finished file is fed back into the checkup. The skill generates a production-prompt file for the PPTX path, and the downstream producer keeps its own mandatory confirmation gates and native export flow, a small but telling respect for the boundary between directing and rendering.

The presentation checkup, and the stage

The checkup is the feature that earns the name humanize. After rendering, it examines the deck for pages that can only be looked at and not spoken, and the repository documents a real case with dates: a June English deck where the static scan passed, but per-page screenshot review found a page-number badge covering body text on nine pages, leaving the audience reading a broken sentence fragment. The checkup flagged it, issued fix instructions, and the recheck passed, without a human counting pages one by one.

Presentation mode completes the delivery loop: in a rendered deck, pressing S switches to a stage view with the current page enlarged beside a timer, the page's speaker script and cues, a preview of the next page, and a full outline. The skill self-produces a basic presenter shell from its slide plan and speaker-intent files, and fuller downstream consoles can take over the stage when the deck renderer provides one.

What the checkup and the stage share is a focus on the speaker rather than the viewer. Most deck tooling optimizes the artifact as a document; this project optimizes the ninety minutes someone stands in front of it, which is a rarer and more defensible obsession.

Evidence, shipped with the deck

The repository argues with artifacts rather than adjectives. The style gallery works by having the downstream renderer genuinely produce four covers, ink classic, kraft paper, indigo porcelain and a Swiss blue, before the outline is fixed, so style is chosen from real output rather than imagination. The visual-enhancement sample set shows a hero image actually generated through the local CLI path and a per-page explanation video genuinely rendered as an MP4, with a dated production record per slot.

Version 1.1 demonstrates the native PPTX path in both languages with two deliberately different sample decks, a Chinese Memphis Pop build and an English Risograph Zine build, built under the same five-page semantic contract, the same design confirmation gate and the same checkup standard, with a bilingual verification record published alongside. The stated point, that one native path can produce genuinely different visual languages, is proven by the artifacts rather than asserted.

This dated-records habit, verification files, production logs, round-by-round checkup notes, is unusual in a skill repository and effectively serves as its own audit trail. It is also, incidentally, how a review like this one can cite specifics instead of vibes.

Structure, licence, limits, and the alternatives

The repository is organized like the product it claims to be: a SKILL.md entry point with agent UI metadata, a specification document, contracts and adapters directories for the downstream relationships, a registry, a versioned Python implementation, tests, and a theory document explaining the audience-state-transition model in depth. Bilingual READMEs serve both audiences.

The licence badge says MIT while GitHub's detection reports no standard licence, so the licence file is worth a glance before redistribution. The limits are structural: quality depends on the downstream rendering skills, the workflow presumes an agent harness with skill support, and the theory, however sound, is one author's model of how talks work. The last push was on 2026-07-31.

The alternatives frame the slot. Using a template deck skill alone gives prettier shells with no line and no checkup, the exact failure this project was built against. Hosted AI deck products produce attractive presentations quickly in the cloud, template-first and subscription-shaped, with less control and your content on someone else's servers. Manual outlining in PowerPoint remains maximal control at maximal labour. Humanize PPT's bet is that the missing layer was the director, and that once the outline thinks like a talk, the renderers already exist.

Editorial conclusion

Humanize PPT fits speakers and trainers already running skill-capable agents who want their material planned as a talk, an audience-state-transition line with per-page intent, real visuals where they belong, and a checkup that catches pages no one could present, in Chinese or English, HTML or native PPTX. It does not fit anyone without an agent harness, or teams expecting a single self-contained deck generator, since rendering is deliberately delegated to downstream skills. Verify first: that your renderer of choice installs and renders on your stack, that the licence file matches the MIT badge given the detection mismatch, and whether your talks match the model's assumptions by reading the theory document before trusting the outline. The last push was on 2026-07-31.

Frequently asked questions

Official sources

1. Issues
2. LearnPrompt/humanize-ppt on GitHub
3. Project website
4. README
5. Releases