codex-ppt-skill: Turn Articles into Image-Based PowerPoint Decks

The article is finished but the talk is tomorrow, and the deck is still a blank title slide. codex-ppt-skill takes the opposite trade from most slide generators: every slide becomes a complete 16:9 image, so the visual style is consistent by construction instead of by an hour of manual restyling.

Why This Skill Matters

Most AI PPT tools hand you editable placeholder slides that you then spend the evening fixing. This skill generates whole-page images — the deck looks like a designer set it, but slide elements are not directly editable. It was built for Codex, where it can use the built-in image generation and editing tools, and it also runs in Claude Code, OpenClaw, Hermes Agent, and other SKILL.md-compatible agents. In those non-Codex environments you usually need to configure gpt-image-2, a third-party image API, or an OpenAI-compatible image endpoint.

The workflow is staged on purpose. Before any full-deck generation, the skill confirms the outline, slide count, visual style, image backend, and one sample slide with you — which is what keeps a ten-page generation from drifting into ten wrong pages.

Installation

The README recommends the skills CLI for a manual install into Codex's global skills directory:

npx -y skills@latest add ningzimu/codex-ppt-skill \
  --skill codex-ppt \
  --agent codex \
  --global

Restart Codex afterwards. For Claude Code, the same command with --agent claude-code installs to ~/.claude/skills/codex-ppt. If you would rather delegate, send your agent one sentence: "Please install this codex-ppt skill for me: https://github.com/ningzimu/codex-ppt-skill".

Real Workflow: Turn a Technical Article into a 10-Slide Deck

1. Point the skill at your source material — articles, reports, papers, course notes, Markdown, outlines, PDFs, and Word documents all work as input:

Use the codex-ppt skill to turn /path/to/article.md into a deck of about 10 slides.

2. The skill reads the content and plans first. It produces outline.md with the slide count, titles, and per-slide key points, and asks you to confirm before generating anything.

3. You choose a visual direction. There are 12 built-in style references — Clean Professional, Scientific Defense, E-ink Magazine, Hand-drawn Technical Explainer, Data Dashboard, McKinsey Style, Teaching Courseware, and more — with 2-3 options offered at this step. The README specifically recommends Hand-drawn Technical Explainer if you do not want to write style prompts.

4. Before the first image, the skill states which image backend it will use and asks for confirmation. Inside Codex with a subscription, the built-in image generation usually needs no API key; third-party OpenAI-compatible endpoints are the fallback path.

5. It generates a single sample slide. Judge typography, layout rhythm, and text quality here — this is the cheapest moment to change course.

6. After you approve the sample, the skill generates the remaining slides page by page, checks text readability, style consistency, and content completeness, writes the speech.md speaker notes, and assembles everything into a .pptx with its local script. Optionally, one subagent handles one slide in parallel after the sample is approved.

Each deck lands in its own project directory:

my-deck/
├── origin_image/     # final image per slide: slide_01.png, slide_02.png, ...
├── outline.md        # the confirmed outline
├── speech.md         # speaker notes, written into the PPT's notes
└── my-deck.pptx      # the assembled deck

A second workflow worth knowing: style replication. Give the agent a screenshot, a PDF, or a PPT/PPTX file whose look you like, and it analyzes the color, layout, typography, and visual elements before generating a new deck in that style. Once you are happy with a deck, ask it to save the style to your personal library at ~/.codex-ppt-skill/references/ — a personal style with the same name takes priority over a built-in one, and the library survives skill updates because it lives outside the install directory.

Tips

• Codex's built-in image generation produces lower-resolution images and does not let you set resolution manually. For sharper text — especially text-heavy slides — switch to the gpt-image-2 API path, which defaults to 2K 16:9 and can go 4K on request.
• Fix one slide, not the deck. If only page 7 bothers you, ask the agent to revise that page in detail instead of regenerating everything.
• Assign real figures to slides. Paper figures, experiment charts, screenshots, and architecture diagrams can be reserved for specific slides in the outline, and the page layout adapts around them.
• The staged flow feels slow on purpose. Confirming outline, style, and one sample slide costs minutes; discovering the style is wrong on page nine costs a full regeneration.

When Not to Use This

Image-based slides are the wrong trade when the deck needs live editing afterwards — you cannot click into a text box on a PNG. The README points to the same author's image-to-editable-ppt-skill for converting generated decks into editable slides; see the image-to-editable-ppt-skill tutorial. And for a low-stakes internal talk where looks do not matter, plain markdown slides are still faster.

See the leaderboard for more skills.