Should AI Stop Outputting Markdown? An Anthropic Engineer Makes the Case for HTML

Right now, virtually every large language model outputs content in Markdown by default. This seems like the obvious choice — LLMs generate text, and Markdown renders headings, lists, bold text, and code blocks using nothing but plain text. It's cheap, universally compatible, and backed by a mature toolchain. What's not to love?

But an engineer on Anthropic's Claude Code team sees things differently. On May 8, 2026, Thariq published a provocatively titled blog post — "Using Claude Code: The Unreasonable Effectiveness of HTML" — arguing that Markdown is fundamentally inadequate, and HTML should become the new default for AI output.

The article quickly crossed a million views. Notable developers like Simon Willison, co-creator of Django, publicly acknowledged it had them "rethinking three years of Markdown-default habits." The debate was on.

"If Humans Aren't Writing It, Why Use a Format Optimized for Humans?"

Thariq's core argument is elegant in its simplicity. Markdown was designed by John Gruber in 2004 with one goal: let humans quickly describe HTML structures using plain text. It's essentially a simplified shorthand for HTML — one that sacrifices enormous expressive power for the sake of easy typing. Side-by-side layouts, interactive elements, embedded SVG charts, color specifications, animations — all impossible in Markdown.

That tradeoff made perfect sense when humans were the ones authoring documents. But when the author is an AI that can generate thousands of tokens per second, does "ease of typing" still matter? What AI needs isn't a format that's fast to write — it's a format that can fully express intent. And HTML, as the native language of the browser, natively supports everything Markdown cannot.

To prove the point, Thariq published 20 single-file HTML examples on his demo site (thariqs.github.io/html-effectiveness), covering code reviews, design systems, prototypes, slide decks, and reports — essentially every scenario a developer encounters daily.

Three Reasons HTML Has a Decisive Edge

Analyzing these examples reveals three core advantages:

1. HTML Preserves the "Dimensionality" of Information

Code diffs, module relationship diagrams, design palettes — these are inherently two-dimensional information. When you cram them into Markdown, you're forced to lay everything out linearly, requiring the reader to mentally reconstruct spatial relationships. HTML lets you render side-by-side columns directly, enabling instant visual comparison.

2. HTML Enables Interactive Understanding

Animation adjustment sliders, clickable prototypes, real-time hash ring demos where you can add or remove nodes on the fly — no amount of static text description can convey these effectively. The difference in communication efficiency between HTML's dynamic visualization and Markdown's flat text is, as Thariq puts it, "orders of magnitude."

3. HTML Is the Content's "Native Language"

Writing a design system's color palette or a component's state variations in Markdown is essentially taking a detour. Instead of listing hex codes in a table, why not render them as actual color swatches? The content should live in HTML — Markdown is just an indirect proxy.

The Practical Constraints That Remain

Thariq's argument, however, applies to a specific scenario: final deliverables generated by AI and consumed by humans. For agent-to-agent data exchange or documents requiring version control, lightweight and easily parseable Markdown (or plain text) is still the right call.

Two real-world challenges stand out:

• Cost: Rich HTML files with inline CSS and JavaScript can balloon to 5–10x the token count of equivalent Markdown. This hits API costs directly.
• Latency: More tokens means longer generation times. In scenarios demanding instant responses, the wait for HTML output degrades the user experience.

Thariq suggests using prompt caching to cache static CSS/JS sections and reduce costs, but this requires sophisticated engineering — not something everyday users can leverage immediately.

The Verdict: Maybe Neither HTML Nor Markdown Is the Answer

Think of Markdown as Excel — limited in capability, but efficient and rock-solid. HTML is more like Google Sheets — far more expressive, but environment-dependent and heavier on performance. These two are likely to coexist by use case rather than one replacing the other.

But zoom out further, and even HTML may not be the destination. It carries 30 years of browser legacy and wasn't designed with AI generation in mind.

The real solution may be the emergence of an AI-native output format — one that delivers HTML-level expressive power at something close to Markdown's generation efficiency. Early signals are already visible:

• Separation of structure and rendering: The model outputs only a semantic structure (like JSON), and a client-side template handles rendering. Claude Artifacts and ChatGPT Canvas are moving in this direction.
• Template-filling approach: AI fills in only the variable portions of a predefined HTML template, keeping token costs low while preserving rich output.

Thariq's core insight holds up: once you free AI from the constraint of "writing fast like a human," limiting its expressiveness feels fundamentally unnatural. Short-term, the answer is pragmatic use-case-driven tool selection. Long-term, we're headed toward an entirely new standard that marries efficient AI generation with rich human-facing presentation — a format that doesn't exist yet, but probably should.