LLM-in-HTML: click any component to fork branching follow-ups instead of chat

Ryoiki-Tokuiten · reddit · 2026-11-20

A Redditor proposes flipping the chat interface: since LLMs are superhuman at one-shot standalone HTML files, put the LLM inside the HTML.

Instead of chat, the model answers by generating an HTML file. Click any component (the app detects it) to ask for a follow-up or deeper version of that section as a branched HTML file; click empty space for a simple follow-up. You can backtrack and branch unlimited times — the demo shows a tree of "branching worlds" fanning out into dozens of standalone pages on astronomy, simulations, and visualizations.

The core insight: HTML is flexible enough to hold almost any output format. Open-sourced as Generative-Recursive-Learning.

Related event: Developer Turns LLM Output into Recursive HTML: Click Any Component to Generate Deeper Layers(3 posts)→

Original post →

More from coding & agent

coding & agent channel →