Using Claude Code: The Unreasonable Effectiveness of HTML

The argument isn't just about aesthetics. Choosing HTML over Markdown when prompting Claude is a structural decision that affects what the model's artifact renderer can actually execute, meaning the output format shapes the model's effective capability ceiling for a given task, not just its appearance.

This sits in an interesting relationship with the 'Quoting Anthropic' piece from Simon Willison earlier this month, which covered Anthropic's own research surfacing gaps between how Claude behaves and how builders assume it behaves. That piece was about alignment blind spots; this one is about output architecture blind spots. Both point to the same underlying issue: practitioners are often working with incomplete mental models of how Claude actually processes and renders information. The HTML-versus-Markdown distinction is a low-cost, high-leverage variable that most prompt engineers haven't systematically tested, which makes it genuinely useful signal rather than theoretical advice.

Watch whether Anthropic formalizes HTML as a recommended output format in its official Claude prompting documentation within the next few months. If it does, that confirms this pattern has moved from researcher observation to supported practice. If it doesn't, the technique remains a community heuristic without institutional backing.