“HTML is the new markdown” (The unreasonable effectiveness of HTML)

Maker: Thariq Shihipar (Claude Code team, Anthropic)
URL: https://twitter.com/trq212/status/2052809885763747935 ; examples at https://thariqs.github.io/html-effectiveness/
Status (10 October 2026): an essay and an example gallery, published around 7 to 8 May 2026 (Simon Willison linked it on 8 May 2026). Widely shared; secondary sources report millions of views within a day, which we couldn’t verify. Discussed further on Lenny’s Newsletter’s “How I AI” podcast. The phrase became Lavish’s tagline.

What it is

An argument, from inside the Claude Code team, that for most non-trivial output agents should write a single self-contained HTML file instead of Markdown, with twenty examples in nine groups: exploration and planning (three code approaches, design directions, an implementation plan), code review (annotated PR, PR write-up, module map), design (living design system, component variants), prototyping (animation sandbox, clickable flow), diagrams (SVG figure sheet, annotated flowchart), an arrow-key slide deck, research and learning (how a feature works, a concept explainer), reports (weekly status, incident timeline) and custom editors (ticket triage board, feature-flag editor, prompt tuner).

The problem it’s solving

Long Markdown doesn’t get read (“trades a document you’d skim for one you’d actually read”), and Markdown flattens spatial information: “diffs and call-graphs are spatial information; markdown flattens them.” Token limits that once justified Markdown no longer bind.

Its path / bet

Make HTML the default output format for anything spatial, interactive or long. It’s a prompting habit, not a product, though it is clearly the thinking behind Claude Code Artifacts a month later.

How it works

Ask the agent for “an HTML artifact” (for example, “Help me review this PR by creating an HTML artifact that describes it”), open the file in a browser, interact, and in the custom-editor examples copy the result back into the session.

Strengths

Names a real shift clearly: producing a UI used to be expensive, so we wrote text; now a purpose-built page per question costs one model turn.
The examples are concrete and teach well. The “custom editing interfaces” group (a triage board whose output you paste back) is the seed of the agent-and-person loop.
Strong on learning: explainers, module maps and annotated flows are exactly where text fails.

Weaknesses / limits

The gallery lists no downsides. Critics raise token cost, security of running generated scripts, models drifting on tag hygiene and nesting, and accessibility.
The page is a broadcast: the round trip back to the agent is copy and paste.
Data is frozen into the page when it’s written.
It doesn’t address where the page runs (a browser) or who can see it, which is why Artifacts and Lavish had to exist.

Relation to fictty

Inspire, and the strongest counter-argument. It’s the best statement of why the terminal is not enough for output, and if it’s simply right, the case for terminal screens shrinks to remote sessions, live data and read-back. Read carefully, it also supports fictty’s idea: most of the twenty examples are a small number of shapes (side-by-side comparison, annotated code, a board, a timeline, a stepper, an editor with an export), which is an argument for a catalogue of components the agent fills, not a fresh page every time.

Could fictty adopt it instead of building?

It’s an idea, not a tool. The practical question is whether fictty should output HTML too. Yes, as an export, not as its primary medium.

What fictty should take from it

Port the gallery. Rebuild the twenty examples as fictty screens. Each one that works is a recipe and a demo; each one that doesn’t marks a real limit of the terminal. It’s the most honest test of the scope we have.
Make “paste the result back” unnecessary. fictty’s read-back and watch turn the editor examples into a closed loop with no clipboard.
Teach with it. The “concept explainer” and “how a feature works” examples are the learning thesis in practice; they should be among the first fictty recipes.

Sources

Thariq Shihipar, post on X and HTML effectiveness gallery
Simon Willison, link post, 8 May 2026
Lenny’s Newsletter, How I AI: HTML is the new Markdown
HTML vs Markdown in Claude Code and Anthropic engineer advocates HTML over Markdown (secondary)

Not verified: the full text of the original post (we read the gallery and secondary summaries, not the post itself), and the view counts.