Humanize PPT

Use this skill when a user wants to turn raw material, notes, voice transcripts, documents, or links into a presentation-ready outline and per-page media decisions before delegating rendering to a downstream skill. Old PPT/PPTX files are not read directly: extract their text first (see scripts/pptx_qa.py's dump/inspect output), then feed that text in as --source. A deck Humanize PPT already rendered goes through --qa-from <rendered.pptx> instead — the presentation checkup, not brief mode.

Positioning

Humanize PPT is a presentation system, born for the talk: an Outline Director (AST audience-state-transfer — every page turn moves the audience forward), a Per-Page Visual-Enhancement Director (real image / SVG diagram / Remotion video), a Production Brief Orchestrator, a Presentation Checkup Runner (演讲体检; formerly the QA loop, CLI flag still --qa-from), and a Presenter-Mode hand-off. The motivation: HTML-PPT template skills are great at concept display but blow a simple idea into a dozen pretty pages, while a real 90-minute talk is ~30 — the pretty shell outruns the content density. Humanize closes that gap: it keeps the beauty (rendered natively by the downstream template skill) and makes it presentable — a line you can stand up and deliver. Downstream template skills own "renders beautifully"; Humanize owns "it's a talk, and someone checked it."

The presentation checkup in one sentence: it does not grade beauty, it grades the outline. It compares every rendered page against its outline page, pulls out the pages that can only be looked at but not spoken from, and keeps going until every page is one the speaker can stand up and present. A failed page, in plain words: a page that holds only a few words and never finishes its point, or a page that fails the audience state transfer it promised (the listener walks out of that page in the same state they walked in). Such a page should not exist; the checkup pulls it out and generates fix instructions.

Humanize is broadly compatible with downstream renderers that can consume plain markdown + JSON. Verified routes are: Chinese HTML → guizang-ppt-skill; English HTML → frontend-slides / beautiful-html-templates; native editable PowerPoint → ppt-master. Other downstreams remain hot-pluggable; support levels live in registry/renderer_registry.json and move only on real output.

It runs before downstream PPT / HTML slide skills and around the post-render presentation checkup. It owns the AST contract, the per-page media decision (does this page need a photo, a system diagram, a 10-second process clip, nothing?), the production brief that the next agent consumes, and the checkup pass on rendered HTML/PPTX. It does not own the rendered deck itself.

There are two human review gates before rendering: the outline preview and the renderer's style gate. HTML routes use Humanize's ≥4 real-cover --style-gallery. PPT Master already owns a mandatory three-stage Confirm UI with native visual previews, so --renderer ppt-master --style-gallery delegates to that gate and writes style_gallery_plan.json + commands/style-gallery/ppt-master-confirm-ui.md rather than duplicating the catalog.

The user calls Humanize PPT once for the brief, hands the brief to a downstream skill for native rendering, then calls Humanize again with --qa-from <rendered.html|native.pptx> to run the 3-iteration presentation checkup. Each iteration writes qa_report.md, fix_prompt.md, and qa_iteration.json. After 3 rounds with remaining failures, status flips to needs-human.

Humanize PPT never copies a downstream skill's template, never injects custom sections into it, and never post-processes rendered HTML or PPTX. Fixes return to the downstream author source. See references/guizang-production-brief-orchestrator.md and adapters/ppt-master-bridge-notes.md.