_{🌐 中文 · English}

Install in 30s

Have your agent (Codex / Claude Code / Hermes …) install Humanize PPT — simplest is to hand it the GitHub link:

Please install the Humanize PPT Skill: https://github.com/LearnPrompt/humanize-ppt

Or one line of npx:

npx skills add LearnPrompt/humanize-ppt -g

Claude Code users can use the plugin marketplace (auto-updates):

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

To run the whole flow, install this set of downstream skills too (Humanize writes the outline and decisions; they render and produce assets):

How to pair the Remotion skills (a bit of hard-won experience): default to remotion-video-production to orchestrate the pipeline, paired with remotion-best-practices while writing code to avoid unstable patterns (misused CSS / Tailwind animation, wrong asset paths); for complex work — captions, charts, 3D, batch templates, automated render pipelines — add remotion-video-toolkit.

One line to the agent: "Install humanize-ppt plus the renderer I need (guizang/frontend-slides/beautiful-html-templates for HTML, or ppt-master for native PowerPoint); add baoyu-image-gen and Remotion only when the deck needs those assets."

Use it in one line

Once installed, run the whole flow with one message — copy-paste it to your agent:

Use humanize-ppt to turn this material into an English presentation deck: produce
the AST outline and per-page intent, render natively with frontend-slides (or
beautiful-html-templates), use baoyu-image-gen for images and remotion for video,
then run the presentation checkup and tell me which pages can't be presented, and
finish with presenter mode.

For Chinese, swap in "Chinese + guizang-ppt-skill". CLI flags, staged control, and re-injection commands live in Advanced usage below — beginners can skip them.

For a PowerPoint deck that stays editable element-by-element, ask Humanize to define the AST, delegate native rendering to ppt-master, then feed the PPTX back into the presentation checkup. Humanize writes ppt-master-production-prompt.md; PPT Master keeps its mandatory confirmation and native export gates.

See it

v1.1: native editable PPTX through PPT Master, in Chinese and English

Chinese showcase	English showcase
Download Chinese PPTX · View source	Download English PPTX · View source

_{▲ Two real PPT Master native exports use the same five-slide semantic contract, the same design-confirmation gate, and the same presentation-checkup standard. Text, shapes, relationship diagrams, speaker notes, and transitions remain in PowerPoint. English has been a full-support path since v1.0; v1.1 adds the native editable PPT Master route rather than reintroducing English support. Read the bilingual verification record.}

Style gallery: downstream renders 4 covers before the outline, so you can pick

_{▲ Four covers of the same deck, rendered natively by guizang-ppt-skill — Ink Classic / Kraft Paper / Indigo Porcelain (Style A) + Swiss Klein Blue (Style B). Humanize emits the spec/command; covers are rendered downstream.}

Visual enhancement: images and video are real output, and they move

_{▲ Left: hero image, generated by baoyu-image-gen via the local Codex CLI (gpt-image, no API key — a low-angle shot with the person layered over a product background). Right: a per-page explainer, a real Remotion-rendered mp4 (shown as a GIF here so you can see it move). Per-slot log: production record.}

_{▲ What they look like inside a deck: the image as a full-bleed cover with a big sans-serif headline over it (left); the explainer video embedded in a content slide (right). The assets are produced downstream; Humanize decides which page needs one, where it goes, and how big.}

Presentation checkup: auto-catch covered text

Before: page badge eats the body text	After: every word is presentable

_{▲ Real case (2026-06-13 English deck): static scan passed, but the per-page screenshot review found a page-number badge covering body text on 9 pages — the audience would read "uires confirmation." Caught automatically, fix prompt emitted, re-check passed — no more hunting page numbers by hand with Codex. Round log · Browse the deck}

Presenter mode: press S in the rendered deck to switch to the stage view

_{▲ In the rendered deck, press S to switch to presenter mode: the current page enlarged plus a timer on the left; the per-page speaker script and cues, a next-page preview, and the full talk outline on the right. Humanize now also writes a baseline presenter-shell.html from slide_plan.json + speaker_intent.md; when the downstream deck is ready, guizang / Neo-Grid style native consoles can still provide the fuller stage view.}

What it solves

I give a fair number of talks. Every time I reached for one of those gorgeous HTML PPT skills I hit the same wall: they're built for concept display — a single idea balloons into a dozen pages, but a 90-minute talk tops out around thirty. The pretty shell runs ahead of the content density; the pages look great and the line doesn't hold.

Humanize PPT fills that gap. It doesn't take over the template's "renders beautifully" job — it turns beautiful into deliverable:

1. AST outline — every page turn teaches the audience something. AST = Audience-State-Transfer. I fed an AI 70+ TED talks and had it distill how a talk carries an audience forward, page by page. Humanize weaves your material into that line: each turn should move the audience to understand a concept, not pile on information.
2. Visual enhancement — the pages that need a picture get a real one. Per page it decides image / SVG diagram / video and hands the plan to downstream: images via baoyu-image-gen (local Codex CLI, no key), video via Remotion, data figures as deterministic SVG. Both the Chinese (guizang) and English paths have image generation wired in.
3. Automatic presentation checkup — stop hunting for the broken page. Text covered by a badge used to mean a back-and-forth with Codex to find which page. Better to let the checkup scan it all at once after render and tell you which page, what's wrong, how to fix.
4. Presenter mode — what ships is something you can take on stage. Human feel, a speaker script, the state transition per page, and the HTML beauty kept intact.

The boundary is clear: the full deck is rendered natively by the downstream renderer; Humanize doesn't copy its template or touch rendered HTML/PPTX. Humanize directs the talk; downstream paints each page.

If you see humanize-ppt-team inside WorkBuddy, treat it as a multi-agent packaging of the same core: AST outline, media decisions, renderer hand-off, presenter mode, and checkup split across specialist agents. It is not a separate monolithic renderer. Chinese positioning note: WorkBuddy Team vs Humanize PPT.

Presentation checkup

First, what a failed page is: one with only a few words that never finishes its point; or one that doesn't complete the audience state transfer it promised, so the listener leaves it in the same state they arrived. Such pages shouldn't exist. HTML's variety is seductive — it's easy to ship a deck where a page says nothing. That's made for looking at, not presenting.

The checkup grades the outline, not the beauty: it diffs each rendered page against its outline page. The failure-mode catalog is in references/qa-failure-modes.md (中文), each mode with "what the audience sees." Things the static scan can't catch (text overflow, badge occlusion, the WebGL-cover static-screenshot trap) are honestly listed as "can't catch yet" and backstopped by screenshot review — leave it empty before staging a fake.

Visual enhancement

Humanize decides per page whether it needs an image / SVG diagram / video and writes it into slide_plan.json's media slots (with asset_path + prompt_hint); the downstream skill produces the real file at that path. The generator is hot-pluggable; recommended:

• image: baoyu-image-gen via the local Codex CLI (--provider codex-cli, uses the logged-in ChatGPT subscription, no OPENAI_API_KEY). Use it for atmospheric / concept / hero visuals — pretty and key-free; keep precise-text or data figures as deterministic SVG (image models garble exact labels).
• video: Remotion renders a duration_s-second deterministic loop (no narration).
• diagram: deterministic inline SVG / HTML, zero dependency, no external call.

v0.9 filled all 8 media slots of one deck with real assets (production record): a real codex hero image + 2 real Remotion mp4s + a real screenshot + deterministic SVGs — proving the slots are real tasks and every generator class plugs in.

Style gallery

Don't make people pick a style blind. HTML renderers use Humanize's ≥4 real-cover gallery. PPT Master already owns a three-stage Confirm UI with native visual-style previews, so --renderer ppt-master --style-gallery delegates to that native gate instead of creating four speculative projects. See references/style-gallery-spec.md and the PPT Master bridge.

Outline preview

Since v0.7 Humanize has its own screenshot-able working draft (not a deck): the audience state-transfer map. Input slide_plan.json, output a zero-dependency HTML page — one row per slide ("page → state the audience walks in with → page intent → state they walk out with") plus a top-line state arc. Five minutes before rendering to spot which page stalls.

Presenter mode

What ships is a talk you can take on stage, not a stack of static pages: per-page speaker script, state transitions, HTML beauty kept. Humanize now writes outputs/presenter/presenter-shell.html directly from slide_plan.json + speaker_intent.md, so you still get a usable presenter shell before any downstream deck exists. Downstream template skills can still provide richer native presenter consoles once the final deck is rendered; Humanize owns "what each page says", the template owns "how the final stage renders."

Advanced usage

python3 scripts/humanize_ppt.py \
  --source examples/01-ai-tool-update/source.md \
  --out .humanize-ppt-runs/ai-tool-update \
  --title "AI 工具更新，不只是功能清单" \
  --renderer guizang --guizang-style A --guizang-theme ink-classic

python3 scripts/humanize_ppt.py \
  --source examples/01-ai-tool-update/source.md \
  --out .humanize-ppt-runs/ai-tool-update-pptx \
  --title "AI Tool Updates Beyond the Changelog" \
  --renderer ppt-master \
  --ppt-master-transition fade

python3 scripts/humanize_ppt.py --source examples/01-ai-tool-update/source.md \
  --out .humanize-ppt-runs/ai-tool-update --title "..." --renderer guizang --style-gallery

python3 scripts/humanize_ppt.py --qa-from <rendered.html> \
  --out <prior out dir> --renderer guizang --guizang-style A --max-qa-iterations 3

# image: local Codex CLI, no key
bun ~/.agents/skills/baoyu-image-gen/scripts/main.ts \
  --prompt "..." --image assets/s01-image.png --provider codex-cli --ar 16:9

# outline preview (audience state-transfer map)
python3 scripts/preview_outline_html.py \
  --slide-plan <out>/slide_plan.json --out <out>/preview-outline.html --title "..."

# demo GIF (record the working drafts into a moving GIF)
python3 scripts/record_demo_gif.py --source examples/01-ai-tool-update/source.md \
  --title "..." --out docs/showcase/demo.gif --covers-dir <real-covers-dir>

What it does

• AST outline: audience, state transfer, page intent, speaking rhythm — every turn moves understanding forward.
• Visual enhancement: per-page image / SVG / video, produced downstream via baoyu-image-gen / Remotion / deterministic SVG.
• Style gallery: ≥4 cover candidates before the outline, each rendered downstream, picked in a zero-dependency gallery.
• Outline preview: the audience state-transfer map from slide_plan.json, reviewed before any render.
• Presentation checkup: per-page outline diff after render, scans failure modes, writes fix prompts, capped at 3 rounds.
• Presenter mode: writes outputs/presenter/presenter-shell.html directly from slide_plan.json + speaker_intent.md — usable standalone before the downstream deck exists; downstream native consoles can take over the full stage view once the deck is ready.

How it differs

In a line: the template renders beautifully; Humanize makes it deliverable, watched, and stage-ready. Upstream and downstream, not competitors.

Verified downstream paths

The Humanize brief is plain markdown + JSON, so it works with both HTML-PPT skills and native PPTX workflows:

English and Chinese are now both full: brief exit works + checkup verified on real output + image generation wired in + renderer-specific failure modes recorded. Guizang has 7 Style A/B rules; the English pair now add 5 static rules for horizontal overflow, low contrast, hyphenation noise, missing font contracts, and missing image alt text. Screenshot review remains part of the process for occlusion, clipping, and visual alignment failures that static HTML cannot prove.

Why AST

• Audience: who's listening, what they know, what they resist.
• State: where they start, where the deck should move them.
• Transfer: how each page drives that transfer.

PPT is not an information container. PPT is an audience state-transfer artifact.

See AST Theory, SPEC.md, v1.1 PPT Master Release Notes, and v1.0 Release Notes.

Safety

• Never copies or post-processes the downstream skill's rendered HTML/PPTX; render problems always become a fix prompt sent back downstream;
• All-local scripts, zero API, zero key (images use the local Codex CLI subscription); no material content leaves the machine;
• The checkup stops at 3 rounds and flags needs-human rather than retrying forever;
• No private paths, accounts, or credentials in the brief or examples.

Reference

• SPEC.md, v1.1 PPT Master Release Notes, PPT Master bridge, Style Gallery Spec, Presentation Checkup Failure Modes, Brief Specification, AST Theory.
• Downstream: guizang-ppt-skill, frontend-slides, beautiful-html-templates, ppt-master, baoyu-image-gen.

License

MIT