Noqta
writing/news/2026/05
● NewsMay 10, 2026·6 min read

Anthropic Engineer Sparks Debate: HTML Is the New Markdown for AI Outputs

Anthropic's Thariq Shihipar argues that HTML, not Markdown, is the right output format for modern AI agents. His essay — paired with 20 self-contained HTML examples — climbed to the top of Hacker News and ignited a wider rethink of how humans consume AI work.

Noqta Team
·EN · FR · AR

Anthropic Engineer Sparks Debate: HTML Is the New Markdown for AI Outputs

A short essay by Thariq Shihipar, an engineer on Anthropic's Claude Code team, has reignited a long-running argument inside the AI tooling community: what file format should AI assistants actually produce? His answer — published on May 8, 2026 and titled "Using Claude Code: The Unreasonable Effectiveness of HTML" — is that, for most non-trivial work, Markdown should be replaced by self-contained HTML files.

The post climbed to the top of Hacker News within a day and drew tens of thousands of reactions on X, where developers debated whether the shift makes sense or simply burns more tokens.

Key Highlights

  • Shihipar argues that AI outputs are increasingly meant to be reviewed, clicked, compared, edited, or shared — work that Markdown can no longer carry on its own.
  • He published a companion site with 20 self-contained HTML artifacts generated by Claude Code, grouped into 9 categories of work where HTML structurally beats Markdown.
  • The categories include exploration and planning, code review, design systems, prototyping, illustrations and diagrams, slide decks, research and learning, reports, and custom editing interfaces.
  • The piece reached the #1 spot on Hacker News on May 9 and triggered active discussion across X and developer Slack channels.

Why The Argument Lands Now

Markdown won the AI-assistant era for a simple reason: tokens were expensive and context windows were tiny. In the GPT-4 era, an 8,192-token cap made every angle bracket a tax you could not afford.

That constraint is gone. Claude, GPT, and Gemini now operate with context windows in the hundreds of thousands or millions of tokens. As one widely shared X post put it: "200k token windows make the tag tax irrelevant and Markdown is a lossy format pretending to be rich text."

Shihipar's argument is that HTML gives the model — and the human reading the output — a much richer canvas:

  • Tables with real headers, alignment, and styling
  • SVG diagrams drawn directly in the response instead of ASCII art
  • Annotated diffs with color-coded severity
  • Interactive widgets: sliders, collapsible sections, in-page navigation
  • Embedded JavaScript for prompt tuners, feature flag editors, or triage boards
  • Color and spatial encoding that Markdown simply cannot represent

What His 20 Examples Cover

The companion site at thariqs.github.io/html-effectiveness ships one HTML file per use case. Among the 20 artifacts:

  • Exploration & planning: code-approach comparisons, visual design directions, implementation plans
  • Code review: annotated pull requests, PR writeups, module maps
  • Design: living design systems, component variants
  • Prototyping: animation sandboxes, clickable user flows
  • Illustrations: SVG figure sheets, annotated flowcharts
  • Decks: single-file slide presentations
  • Research: feature explainers and concept explainers with in-page navigation
  • Reports: status reports and incident timelines
  • Custom editing UIs: triage boards, feature flag editors, prompt tuners

Each artifact is a single .html file that opens in any browser — no build step, no framework, no server.

The Pushback

Not everyone is convinced. The replies under Shihipar's announcement split roughly three ways.

The agent-first camp argues that Markdown is still the right format when one AI model is producing output for another. Lower token cost, easier parsing, deterministic structure. "Markdown for agents, HTML for humans," one widely upvoted reply summarized.

The practicality camp notes that GitHub gists, Notion, Linear, and most chat tools render Markdown beautifully out of the box. An HTML file requires opening a tab, hosting somewhere, or embedding via iframe.

The token-economy camp is more cynical. "Anthropic definitely wants us to spend more tokens on HTML instead of Markdown," one developer wrote, pointing out that richer outputs are also more expensive outputs.

Impact On Tools And Workflows

Within hours of publication, a community-built Claude skill called html-artifacts appeared on GitHub, encoding Shihipar's recommendations as a reusable instruction set. The skill tells Claude to produce a self-contained HTML artifact whenever the task involves anything spatial, interactive, or visual — and to fall back to Markdown only for plain prose or short answers.

The conversation also lands at a moment when Anthropic, OpenAI, and Cursor are all shipping more "artifact-style" features — interactive HTML rendered directly in the chat surface. ChatGPT recently expanded its Canvas mode, Claude's Artifacts panel has become a daily-driver workflow for many developers, and Cursor's workspace now treats agent-generated HTML as a first-class deliverable.

What's Next

Shihipar's essay does not announce a product change. It is a personal blog post from one engineer. But it formalizes a pattern many Claude Code users had already drifted into, and the SKILL.md open standard adopted by 30+ AI coding tools earlier this year makes it trivial to distribute an "HTML-first" preset to entire teams.

The likely next step is platform support: editor extensions that auto-render agent HTML, sandboxed previews inside chat clients, and conventions for when to attach a one-page .html versus when a Markdown reply is enough. The bigger shift Shihipar names — that AI outputs are no longer documents but temporary thinking environments — is the part the industry will be working through for the rest of 2026.

Source: Using Claude Code: The Unreasonable Effectiveness of HTML — companion site and Simon Willison's coverage