HTML Over Markdown

For most of the current generation of AI tooling, the default output format has been Markdown. It is what chat interfaces render, what coding agents emit, and what every frontier model has been tuned to reach for. It is also, increasingly, the wrong default.

The reframe — stated publicly by Andrej Karpathy and arrived at independently by practitioners on the Claude Code team such as Thariq — is simple: Markdown is a report; HTML is an interface. A report is something you read. An interface is something you act on. As models get better at producing rich, self-contained output in a single shot, the format you ask them to use decides whether the recipient receives a wall of text or a small working application.

This matters most inside enterprises, where the bulk of AI output is deliverables — analyses, status documents, dashboards, reviews — that someone is supposed to act on. Delivered as Markdown, they are read and set aside. Delivered as self-contained interactive HTML, they become artifacts the recipient can filter, manipulate, and continue working inside.

This is not a universal upgrade. HTML is heavier, harder to version-control, and carries real security and accessibility costs when produced carelessly. Markdown remains the correct choice for source-controlled documents and for any text destined to be read by another model.

This paper argues that output format is a user-experience decision, not a default — and lays out the patterns Tenten AI uses to make that call, and to turn HTML into a reliable output mode for production coding agents.