HTML slides without PowerPoint: a 2026 workflow with Codex + GPT-Image-2

PowerPoint is shaped like 1995 — a file format, a fixed canvas, and an editor that fights you the moment you want to embed code or live data. The replacement most engineers settle on is HTML: one file, opens in any browser, version-controllable, embeddable in docs. What changed in 2026 is that the painful parts — writing the layout, generating consistent figures, dealing with non-Latin text in image generation — are now handled by Codex and GPT-Image-2 directly. About 30 minutes per deck once you've done it once.

Four HTML-slide stacks worth knowing

Pick whichever maps to how you already think.

1. Slidev — Vue-based, slides written in Markdown with --- separators. Native code blocks with syntax highlighting, line-by-line reveals, live components. Exports to HTML, PDF, SPA, or PPTX. Pick this if your decks have any code in them.

2. reveal.js — The original. Plain HTML/CSS/JS with a thin slide framework on top. Heavy plugin ecosystem. Pick this if you want maximum control and don't mind writing HTML directly.

3. Marp — Markdown-only, no JS framework, opinionated themes. VS Code extension previews live. Pick this if you want "type Markdown, get a deck" with zero ceremony.

4. html-ppt-skill — Not a framework — a Claude/Codex agent skill that generates standalone HTML decks with 24 themes and 31 layouts pre-baked. The agent picks the layout per slide based on content. Pick this if you want the AI to handle layout choices, not just content.

A common pick is Slidev for technical talks (code-heavy decks) and html-ppt-skill for business decks (where layout variety matters more than code).

What changed: GPT-Image-2

The bottleneck on HTML slides used to be figures. GPT-Image-2 (released 2026-04-21) fixed two things specifically relevant to slides:

1. Multilingual text accuracy, including Chinese, Japanese, Korean — at the character level. You can now generate a chart with Chinese labels that aren't garbled.
2. Agentic layout reasoning — the model plans the figure structure before rendering, so you get fewer "almost right" iterations.

Codex can call GPT-Image-2 directly from a session — you describe the figure in plain language, the image arrives back, you reference its file path in the slide HTML.

The recipe

Prereqs: Codex CLI installed and access to GPT-Image-2 through Codex (the standard ChatGPT Plus / Pro plan, or any provider that proxies the same APIs). Your outline in any form: bullets, Notion doc, transcript. About 30 minutes once you've done it once.

Step 1 — Hand the outline to Codex. In a Codex session:

Here's an outline for a 12-minute talk on <topic>. Generate it as a Slidev deck. One slide per major bullet. For each section, suggest where a figure would help, but leave a placeholder — I'll fill those in step 2.

Read what comes back. At this stage check structure, not polish — does slide order match how you'd actually talk?

Step 2 — Generate figures in the same session. For each placeholder:

For slide N, generate a figure with GPT-Image-2: <description>. Use the cover art's palette (cream and dark blue). Any text in the figure should be in Chinese.

Two things matter: (1) specify the palette explicitly — defaults trend saturated; (2) specify the language of text in the image — without it you'll get English even in a Chinese deck. If a figure isn't right, ask for a variation. About $0.03–0.05 per image at typical output sizes; iterating is cheaper than fighting.

Step 3 — Theme + content pass.

Run the deck through Slidev locally, screenshot the first three slides. What looks off?

Common fixes: title overflow → shorter title; image bleeds outside slide → adjust aspect ratio; code unreadable → bump font or split slide.

Step 4 — Present + export. slidev dev for presenting (presenter view at ?presenter). slidev export for PDF, slidev build for a static SPA, slidev export --format pptx if the recipient lives in PowerPoint (animations don't survive the conversion).

Variations

• Conference talk — Slidev with --listen for remote presenter notes from your phone
• Board / investor deck — html-ppt-skill, ask for "executive theme", export to PDF
• Tutorial deck for a recorded video — reveal.js with the chalkboard plugin so you can draw on slides
• Dataviz-heavy deck — Slidev with Observable plot embeds; Codex writes the plot specs from your CSV

What this workflow does not do well

• Agency-grade brand decks — Figma + Keynote will be faster if margins matter to a brand team
• Heavy animation — click-to-reveal is clean; "fly in from the left" is more friction than it's worth
• Collaborative editing — it's a Markdown file in git; if your team works in Google Slides for real-time co-editing, wrong fit
• One-off throwaway decks — Gamma + 6 bullets is faster