Markdown Can't Do This (HTML Can)

A 58-second Short, and a trailer for the channel’s longer video. It reports that Thariq Shihipar of Anthropic’s Claude Code team has argued that Markdown is the wrong format for the specs, plans and pull-request write-ups people trade with coding agents, and that HTML should replace it. The Short’s case fits in two lines. Markdown carries “headings, bold, a bullet list” and nothing more; HTML carries tables, styling, SVG charts, interaction and real images, so an agent asked for HTML “stops faking the chart and just draws it”. It closes on the claim that a colleague will skim Markdown and open HTML.

The post behind it is “Using Claude Code: The Unreasonable Effectiveness of HTML”, published on X on 8 May 2026 and republished on Anthropic’s blog on the 20th as his personal opinion. It is narrower than the Short. Shihipar is talking about what an agent hands back to a person: he says he tends not to read a Markdown file past a hundred lines and cannot get colleagues to, and that HTML keeps him “much more in the loop”. He calls himself an HTML maximalist. He does not say Markdown is dying; that is the Short’s headline.

Two of the Short’s details are wrong. Markdown does more than it says: CommonMark specifies code blocks, links, images, block quotes and raw HTML, and GitHub’s flavour adds tables. What it lacks is styling, charts and interaction, which is the real argument. The post went up the day before the Short, not “last week”. Its “750,000 views” may be an early snapshot; one report put the post at 4.4 million in its first 16 hours.

The Short leaves the cost for the long video. By secondary accounts the original post put HTML at two to four times slower to generate. Corey Gallon measured 1.4 to 2.4 times, and found that the larger cost comes when an agent has to read its own HTML back in. Kurtis Redux’s reply argues from diffs, version control and security. Simon Willison, often listed as a critic, wrote that the post made him reconsider.