English · 简体中文
Codex PPT is a Codex skill for creating polished 16:9 presentation decks in two modes: image mode, and editable PPTX mode that first generates approved slide images and then reconstructs each slide into editable PowerPoint text and movable visual layers.
It is built for the kind of deck where the story, visual system, slide prompts, QA loop, and final export all matter. Instead of asking an AI model for "some slides", the skill makes Codex behave like a deck producer: plan first, lock the visual language, generate each slide as a complete image, reject weak pages, and then either compile the image deck or run an editable reconstruction pass.
- Added
editable-pptxmode: generate full-slide images first, then reconstruct each page into editable PowerPoint. - Kept default
imagemode: generate slide images and optionally package them as PDF. - Added a per-slide subagent contract for visual layers,
text-layer.json, and QA notes. - Added editable PPTX build and merge scripts to assemble per-slide outputs into one final
.pptx.
- Turns a topic, outline, report, or rough notes into a complete slide deck plan.
- Writes a brief, storyline, design system, storyboard, and one prompt per slide.
- Generates every final slide as one full-page AI image, including layout, text, charts, and visual objects.
- Forces QA before delivery: aspect ratio, legibility, page order, style consistency, and text accuracy.
- Compiles approved slide images into a strict 16:9 PDF.
- In editable PPTX mode, runs one reconstruction worker per slide and merges the outputs into a final editable
.pptx.
image: default. Generate full-page slide images; compile a PDF unless the user only wants image files.editable-pptx: generate and QA the images first, then reconstruct each page into native editable text boxes and movable visual layers before exporting.pptx.
The key idea is simple: in image mode, if a slide fails QA, rewrite the prompt and regenerate the whole slide. Do not locally patch text or layout. Editable mode starts only after the image slides pass QA and keeps derived artifacts under editable/.
Clone this repository into your Codex skills directory:
git clone https://github.com/qybaihe/codex-ppt.git ~/.codex/skills/codex-pptThen ask Codex:
Use $codex-ppt to make a 10-page product pitch deck about my app idea.
Local script dependencies:
pip install -r ~/.codex/skills/codex-ppt/requirements.txt
npm install --prefix ~/.codex/skills/codex-pptFor editable output:
Use $codex-ppt to make an editable PPTX deck about my app idea.
The repository includes a real generated deck:
.
├── SKILL.md
├── agents/openai.yaml
├── references/slide-quality-checklist.md
├── references/editable-pptx-mode.md
├── references/editable-slide-subagent-prompt.md
├── scripts/compile_slide_images_to_pdf.py
├── scripts/build_editable_ppt_from_layers.mjs
├── scripts/merge_editable_slide_outputs.py
├── assets/
└── examples/shit-app-memphis-pitch/
Codex PPT deliberately separates deck production into phases:
- Brief: define audience, objective, page count, tone, source material, and constraints.
- Storyline: create a narrative chain before designing pages.
- Design system: lock palette, type hierarchy, grids, slide variants, image language, and anti-patterns.
- Storyboard: specify each page's role, title, takeaway, copy plan, layout, visual content, and QA risks.
- Generation: create one prompt per slide and generate full-page slide images.
- QA: reject malformed text, style drift, clutter, cropped content, and weak covers or endings.
- Export: compile approved images into a 16:9 PDF, or continue into editable PPTX reconstruction.
Many AI PPT attempts fail because they stitch together generic templates, local text overlays, and inconsistent visuals. This skill takes the opposite route: each slide is one coherent generated poster-like page. That makes it especially useful for pitch decks, concept decks, launch decks, visual reports, mood-driven presentations, and experimental product storytelling. When downstream editing matters, editable PPTX mode uses the approved image deck as the visual target and rebuilds PowerPoint-native text and movable layers.
MIT