Ask AI to write a report and it always comes back as Markdown. Lines of ##, bold text, and bullet points. You read the first few lines, then scroll faster and faster. Past 100 lines? Honestly, you just close it.
Thariq Shihipar from Anthropic's Claude Code team dropped a line on X in May that hit 10.9 million views in 48 hours: "HTML is the new Markdown."
So what's wrong with Markdown?
Markdown becoming the default AI output format is a relic of the GPT-4 era. Back then, context windows were 8,192 tokens. HTML uses more tokens structurally, so Markdown made sense in tight spaces.
That habit persisted into 2026 — but the situation has completely changed. Context windows are now at 1 million tokens. Even if HTML costs 5,000 tokens more, that's just 0.5% of your total budget.
And there's a more fundamental shift. Thariq's core observation: "I don't manually edit AI-generated documents anymore. I use them as specs or pass them into the next Claude session." Markdown's biggest advantage was being easy to edit by hand — but that premise just collapsed.
Key stat
According to Thariq, about 99% of AI-generated tokens aren't production code. They're dashboards, interfaces, planning documents, status updates. The shift toward visualization is already happening.
What actually changes?
Try it yourself and you'll feel it immediately. Simon Willison analyzed a security exploit using this prompt: "Explain this code in detail. Expand out any confusing bits. Output HTML, neatly styled and using capabilities of HTML, CSS, and JavaScript to make the explanation rich and interactive."
The output came with a high-level summary, safety warnings, step-by-step code analysis, and comparison tables — all with tabs and collapsible sections. Stuff that's really hard to pull off with equivalent clarity in Markdown.
| Markdown | HTML | |
|---|---|---|
| Readability (100+ lines) | Eyes glaze over | Tabs, collapsibles, side nav |
| Diagrams | ASCII art | Inline SVG |
| Interaction | Not possible | Sliders, buttons, toggles |
| Comparison tables | Pipe characters | Sort, color, highlight |
| Sharing | Needs a Markdown viewer | Opens in any browser |
| Generation speed | Fast (30s) | 2–4x slower (90–120s) |
| Git version control | Clean diffs | Noisy diffs |
The core HTML advantage: "There is almost no set of information that Claude can read that you cannot fairly efficiently represent with HTML." That's Thariq's own phrasing. Add SVG diagrams, CSS, and JavaScript — and you get output people actually want to read.
That said, HTML isn't always the answer. Markdown still wins for Git repos (diffs get noisy), agent-to-agent handoffs (needs clean parsing), short replies under 20 lines, and personal notes or drafts.
How to start: the essentials
- Add one line to your prompt
Append to any existing prompt: "Output as a single self-contained HTML file." That alone gets Claude to add structure and styling automatically. - Use it for finished deliverables only
Keep drafts and iteration in Markdown. Switch to HTML for final reports, client work, and anything meant to be shared. No need to flip everything overnight. - Tell it the purpose
"A human will read this. Add tab navigation. Make it print-friendly." Or: "Let me adjust parameters with sliders." The more context you give about how you'll use it, the better the result. - Give it creative freedom
Thariq's tip: the phrase "whatever is needed" is surprisingly powerful. "SVG, code snippets, mockups — whatever gives me maximum context." Trusting the model unlocks more creative output. - Try a Living Design System
Build a single HTML file with your project's colors, typography, and components. It's a shared reference that both humans and AI can read and use.
