2026-05-12 HTML as the New AI Output Format

Canonical version: 2026-05-12 HTML as the New AI Output Format.

HTML is another output format for AI-generated content, next to markdown. It can include animations, interactivity, charts, layout, and structure that text alone can't carry. Markdown remains the better fit for text-first content.

This week, Thariq Shihipar (Claude Code team) published a long post on HTML as an AI output format, and Andrej Karpathy placed it on a broader spectrum of how humans and AIs exchange information.

What HTML can show

Markdown is great for notes, drafts, and transient conversation. It can't render a working visualization of an algorithm, a slider to tune a design parameter, tabs in a long spec, hover-revealed detail in a report, or an SVG diagram of a system. HTML can.

What HTML adds beyond markdown:

• Tables with proper structure and CSS styling
• SVG diagrams, illustrations, flowcharts
• Animations and transitions
• Interactive widgets (sliders, toggles, buttons) via JavaScript
• Spatial layouts (absolute positioning, canvases)
• Mobile-responsive structure for longer documents

Karpathy's Framing

Karpathy describes AI output formats as a spectrum, where each format keeps its uses:

1. Raw text → 2. Markdown → 3. HTML → 4. Interactive neural videos/simulations

According to him, audio is the preferred human input to AI; richer visual formats are increasingly the preferred AI output to humans. Roughly one-third of the human brain is dedicated to visual processing, so ideas that are better seen than read benefit from the richer formats.

Practical Impact

• To get HTML, ask your AI to "Structure your response as HTML"
• Good fits: specs, code reviews with annotated diffs, design prototypes, research reports, custom editing UIs, interactive explainers
• Always include an export button ("copy as JSON / copy as prompt") to bring interactive changes back into the AI conversation
• Tradeoff: HTML takes 2–4× longer to generate than markdown, and HTML diffs are noisier in version control, so use it when the visual/interactive payoff justifies the cost

Connection to Notes and Wikis

This pairs naturally with LLM Wiki and vault notes: feed a set of permanent notes or wiki articles into a prompt and ask for an HTML page or site. You get a shareable, interactive page built from your notes, without publishing infrastructure, and the markdown notes stay as they are.

See also: Structuring AI responses as HTML

References

• trq212 — HTML as AI output format
• Andrej Karpathy — vision as preferred AI output

Related

• Structuring AI responses as HTML
• LLM Wiki

About Sébastien

Ready to get to the next level?