Claude Code team promotes HTML to replace Markdown: AI output format is being redefined

Core conclusion

On May 8-9, 2026, Anthropic Claude Code team member Thariq Shihipar published a long article on X, systematically demonstrating the advantages of using HTML to replace Markdown as the AI output format. Simon Willison, a benchmark figure in the Python ecosystem, immediately followed up to verify and make the switch public. This marks that the AI ​​collaboration output format is undergoing an important paradigm shift - from "token-saving plain text" to "rich interactive visual output".

Key Points

• Event: Claude Code team member Thariq Shihipar posted a post suggesting that developers require Claude to output HTML instead of Markdown
• Influenced people: All developers using Claude Code / AI coding tools
• Core reason: HTML provides interactive SVG charts, inline navigation, data visualization and other capabilities that Markdown cannot achieve.
• Simon Willison Verification: Use GPT-5.5 to create an HTML version of the vulnerability analysis page, which is far more effective than Markdown
• Practical Suggestions: Use HTML for one-time analysis/interpretation, and Markdown for documents that require collaborative iteration.

Background: From Token restriction to output freedom

In the GPT-4 era (2023-2024), the 8,192 token limit makes Markdown the only reasonable choice. Markdown is more token-less than HTML and can convey more information in a limited context window. At that time, if you asked GPT-4 to output HTML, you would probably waste half the context window on the label.

However, new generation models such as Claude Code and GPT-5.5 have a large context window of 100K-200K, and token efficiency is no longer a bottleneck. This made developers start to re-examine: Since the output space is sufficient, why not pursue better presentation effects?

Thariq made a key insight: the quality of the HTML generated by AI is surprisingly high - the model's training corpus contains a large proportion of HTML, and its ability to understand and generate HTML is far beyond what many people imagine.

HTML vs Markdown: Key differences

Practical combat: HTML output in Claude Code

Thariq recommends using prompts like this in Claude Code:

Help me review this PR by creating an HTML artifact that describes it. I'm not very familiar with the streaming/backpressure logic so focus on that. Render the actual diff with inline margin annotations, color-code findings by severity and whatever else might be needed to convey the concept well.

The effect is that Claude Code will render an interactive HTML page directly in the terminal, including:

• Code diff with inline comments
• Color-coded issue classification by severity
• Clickable chapter navigation
• Embedded SVG flowchart

Simon Willison’s practical verification

Simon Willison uses GPT-5.5 to analyze the Linux kernel vulnerability PoC on copy.fail. The prompt is as follows:

Explain this code in detail. Reformat it, expand out any confusing bits and go deep into what it does and how it works. Output HTML, neatly styled and using capabilities of HTML and CSS and JavaScript to make the explanation rich and interactive and as clear as possible.

The result is a single page of HTML, which contains formatted and rearranged vulnerability code, line-by-line annotations, and interactive charts to illustrate the vulnerability principle. It does not require any external dependencies and can be viewed immediately after opening a file. Simon wrote in his blog:

"I have used Markdown for output by default since the GPT-4 era. This article by Thariq made me rethink, especially the output scenario."

Community Discussion

The thread received 75+ likes and 32 comments on HN, with community reaction being polarizing but generally positive.

Proponents believe this is an underrated practice. Developer shared: "I've been using Claude for the past month to create interactive HTML widgets - much more fun than making slideshows and static documents." Another developer said he has used it for PR Review Dashboards, Workshop interactive pages, and internal memos.

The challenger points out three key trade-offs:

1. Token cost is higher: HTML consumes 3-5 times more tokens than Markdown, which is a business benefit for Anthropic
2. Reduced human editability: If you need to manually modify the AI output, Markdown is far easier than HTML
3. Feedback Accuracy: Annotating modification requests in HTML is much more difficult than in Markdown

Practical Advice: How to Get Started

For Claude Code users, here’s what you can try now:

1. The next time you ask Claude to explain your code, add "Output as HTML" to the end of the prompt
2. Open the generated HTML file directly with a browser (no server required)
3. Explore more uses: PR Review, design documents, data analysis reports, interactive tutorials

For other AI tool users, models such as GPT-5.5 and Gemini are also good at generating HTML based on the same principle.

Tool entry

Claude, Claude Code, OpenAI, ChatGPT, GPT-5.5, Anthropic, Hermes Agent

Internal link guidance

• Want to learn more about Claude Code? Watch the tutorial: Claude Code automated writing practice: build an AI content production pipeline in 30 minutes
• Someone made money using Claude Code: Claude Code 48 hours to start a business: one person + US$29 monthly fee, monthly income in 3 months $9,000
• Use n8n to build a complete automated assembly line: A real case of a data analyst using Claude Code + n8n to build an automated report SaaS with a monthly income of $3,800