HTML Is the New Markdown: Inside the Shift Anthropic's Claude Code Team Just Triggered

In May 2026, a Claude Code engineer argued the AI ecosystem has been quietly anchored to the wrong default output format. The reaction online was instant — and the implications run all the way to how we share work.

The viral moment

On May 8, 2026, Thariq Shihipar, an engineer on Anthropic's Claude Code team, published an essay titled "Using Claude Code: The Unreasonable Effectiveness of HTML." Within 48 hours, the post hit hundreds of thousands of views on X, sparked thousands of quote-tweets, and forced a question the AI development community had been quietly avoiding for four years:

Why is Markdown still the default output format for frontier models?

The argument is short. It is sharp. And the moment Simon Willison wrote it up on his blog, it went from "interesting take" to "industry-wide rethink."

The core argument in one sentence

If your reader is a model, use Markdown. If your reader is a human, use HTML.

Markdown won by default in the GPT-4 era because every token mattered. The 8,192-token context window made any whitespace-friendly, syntax-light format genuinely worth optimizing for. A pretty HTML report could blow your entire context budget on <div> tags.

Four years later, Claude Opus 4.7 ships with a 1,000,000-token context window. A 2,800-token Markdown document expands to roughly 8,000–11,000 tokens as HTML — and that's still 0.5% of the budget. The constraint that made Markdown necessary is gone. The default it created survived anyway, unexamined, for nearly four years.

What HTML lets you do that Markdown can't

The Anthropic engineering argument isn't just about aesthetics. It's about information density and human evaluation:

Markdown gives you headers, lists, code blocks, and bold. HTML gives you the entire web platform.

The four artifact categories Thariq named

In his work and his subsequent interview on Lenny's Newsletter ("How I AI: HTML is the new Markdown"), Thariq highlighted four recurring use cases where his team replaced Markdown with HTML artifacts:

1. Project plans and specs

Plans used to be Markdown checklists. Now they're HTML files with embedded mockups, code excerpts, severity color-coding, and clickable sections. A plan you'd skim in Markdown becomes a plan you'd actually read in HTML.

2. Throwaway micro-apps

A custom UI for editing a specific table of rules. A gamified interface for ranking design tokens. Tiny, single-purpose applications that exist for one task, then get discarded. Generating these by hand would never have been worth it. Generating them with Claude takes 30 seconds.

3. Living design systems

A single self-contained HTML file documenting colors, typography, spacing, components, and motion. Humans browse it like a styleguide. Agents read it like a spec. Both audiences are served by the same artifact.

4. Status updates and reports

Weekly updates that used to be plaintext or Markdown become HTML pages. They get read because they look read-worthy. They get acted on because the visual hierarchy actually signals priority.

The "compute allocator" mindset

Thariq's broader framing in the Lenny's interview reframes the job of an engineer in the age of frontier-model agents: engineers are compute allocators, not code authors. Only about 1% of the tokens an agent generates ever reach production. The other 99% improves the development process — better plans, clearer specs, richer review surfaces, more legible decisions.

HTML artifacts are for the 99%. They're the cognitive scaffolding around the actual code change. And because LLMs can produce them cheaply, you can have one per plan, per PR, per status update, per meeting.

The token economics, made concrete

For anyone who still cares about cost — and you should, but in proportion — here's the math:

FormatTypical tokens for the same content% of 1M-token budget
Markdown2,8000.28%
HTML8,000–11,0000.8–1.1%

The cost difference at the token level is real. The cost difference relative to the value of a human actually engaging with the output is negligible. Markdown was rational when models could only see 8K tokens. It's an outdated default now.

What this means for tooling

If HTML is the new default output, two things have to be easy:

1. Generating HTML from agents — already solved. Claude, Cursor, Windsurf, GPT all produce excellent HTML on demand.

2. Publishing HTML so humans can actually see it — historically not solved. AI generates a beautiful page, then it lives on your disk forever because the publish step is too annoying.

That second problem is exactly the gap host-html was built to close. Your agent generates an HTML artifact. The agent calls one API. You get a URL. You paste it in Slack, email it to a client, project it in a meeting. The "publish" step is one sentence — "publish this as a web page" — not a deploy pipeline.

The publishing primitive

Concretely, here's the simplest possible loop with host-html:

HOSTHTML_API_KEY="hh_YOUR_API_KEY"
curl -X POST https://suifgsvtcbrawzdhyuhk.supabase.co/functions/v1/publish \
  -H "Authorization: Bearer $HOSTHTML_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"html": "<h1>Hello from a Claude HTML artifact</h1>"}'

You get back a JSON payload with a public URL. Your agent can do this autonomously. Generate a personal API key in your host-html account (free) and pass it as the bearer token.

Or, if you're using Claude Code or another skills-protocol-aware agent, install the publishing skill:

npx skills add phanosh/host-html --skill hosthtml-publish -g -y

Then in your conversation, just type "publish this as a web page" and the agent handles it.

The contrarian read

Not everyone agrees with the HTML-is-the-new-Markdown framing. Two pushbacks worth taking seriously:

1. Token cost still matters in batch and tool-call settings. Inside tool outputs that feed back into the model, Markdown remains the right call. The HTML shift is specifically for final outputs intended for human consumption, not intermediate agent state.

2. HTML can hide intent. A Markdown doc forces you to think structurally. A pretty HTML doc can disguise weak reasoning behind nice typography. Use HTML for delivery, Markdown for thinking.

Both critiques are correct. Neither contradicts the central point: for outputs humans actually read, HTML is the better tool.

Where this is heading

If the default flips — and the evidence over the last two weeks suggests it's flipping faster than anyone expected — the next year of AI tooling looks different:

Two months ago, "AI agent that publishes HTML to a public URL" sounded like a quirky workflow. Today, after Thariq's article, it's becoming the default endpoint for any artifact intended for a human.

How to start

The fastest possible way to feel the difference is to try it once:

1. Open Claude Code (or your preferred agent)

2. Ask: "Build me a single self-contained HTML file that explains [topic you care about] with embedded SVG diagrams, color-coded sections, and a sortable table for the supporting data"

3. Save the result, or drop it on host-html, or have your agent publish it directly

4. Share the URL

The first time you do this you'll understand viscerally what Thariq's argument was about. Markdown gives you a document. HTML gives you a page — and pages are how humans actually navigate ideas.

Related reading

Sources

Publish your first HTML artifact in 5 seconds →