For years, developers have reflexively asked AI assistants to "output Markdown" — as if compact syntax and token efficiency were sacred cows. That orthodoxy is now being challenged from inside Anthropic, and the debate reveals something deeper: AI output format has become a first-class product philosophy question, not a mere engineering detail.
Thariq Shihipar, a member of Anthropic's Claude Code team, published what Simon Willison called "The Unreasonable Effectiveness of HTML" — a case against the Markdown conventional wisdom. The argument is straightforward: when you ask an AI for output in HTML rather than Markdown, you unlock SVG diagrams, interactive widgets, in-page navigation, and rich visual explanations that plain text formatting simply cannot express.
This matters enormously for developer tooling. Consider a code review prompt: "Create an HTML artifact that describes this PR, focusing on the streaming/backpressure logic. Render the diff with inline margin annotations and color-code findings by severity." In Markdown, you're stuck with static text. In HTML, Claude can produce something navigable, visual, and genuinely useful.
The tension runs deeper than format preference. Markdown advocates have long argued for token efficiency — the format's lightweight syntax meant more room for actual content within context windows. During the GPT-4 era with its 8,192-token limits, this was a legitimate concern. But as context windows expanded and models became more capable, the calculus shifted. "I've been defaulting to asking for most things in Markdown since the GPT-4 days," Willison admitted — before conceding that Thariq's argument had caused him to reconsider.
Yet not everyone is convinced. Some developers argue that HTML output introduces fragility — browser rendering inconsistencies, CSS dependencies, and the cognitive load of debugging markup rather than content. Others worry about prompt complexity: getting rich HTML often requires explicit instruction, whereas Markdown's formatting emerges naturally from clear requests.
The real stakes became visible when Willison tested the approach with GPT-5.5 on a recently disclosed Linux security exploit. The result was a dark-themed, syntax-highlighted technical document with structural navigation — something Markdown cannot approximate regardless of prompt engineering. It wasn't just prettier. It was meaningfully more comprehensible.
This debate signals a broader shift: AI output format is no longer about bandwidth conservation. It's about semantic richness — what the format can express, not just how efficiently it compresses. For developer tools specifically, where precision and navigability directly impact productivity, the question "Markdown or HTML?" is really asking: "What kind of cognitive artifact do you want the AI to build for you?"
The answer Anthropic's team is pushing toward: something you can actually use, not just read.