Anthropic's Claude Code team member Thariq has fully abandoned Markdown. Instead, he defaults to having AI generate HTML files. His argument: as AI-generated plans get more complex, HTML's information density and bidirectional interaction blow plain text out of the water.
https://twitter.com/trq212/status/2052809885763747935
Thariq points out that Markdown over 100 lines is a readability nightmare. It also forces large models to draw diagrams with garbled ASCII characters. With HTML, you can use CSS and SVG to visually present architecture diagrams and interface designs — and even build deeper interactivity.
His killer use case: temporary editing interfaces. When he can't easily describe a requirement in words, he asks Claude to generate a web page with draggable cards or color sliders, plus a "Copy as JSON" button. After tweaking the result intuitively in the browser, he copies the parameters back to the command line and feeds them to the model. The loop closes. It’s a hands-on workflow that fits with Anthropic’s vision of more interactive AI, like its experimental Orbit assistant for Gmail and Slack.
Yes, it takes 2 to 4 times longer for the model to write HTML than Markdown. Yes, diffing in version control is harder. But Thariq says the experience of being pulled back into the decision loop is totally worth it. And the company recently open-sourced a natural language autoencoder to make AI's 'inner thoughts' legible — suggesting a broader trend toward richer outputs.