Back to Podcast Digest
AI Engineer7m

HTML is All You Need (for Agents to Make Graphics) - Amol Kapoor, Nori

TL;DR

  • The medium is the problem, not the model: Agents fail at SVG and canvas tools not because they lack spatial reasoning, but because SVG is a wall of numbers that neither humans nor models can work with intuitively.

  • HTML is the native language for AI graphics: Models have trained on billions of HTML examples and understand its structural semantics (headings, grids, charts) intuitively, letting browsers handle pixel rendering.

  • The pelican test exposes the gap: Developer Simon Willison asks models to draw a pelican riding a bicycle using SVG, with consistently terrible results. The same task in HTML works because models reason about structure, not coordinates.

  • PowerPoint is just a tool, not the deck: The presentation mode is what audiences see, and the editing format is arbitrary, so you can simply choose HTML as your editing format.

  • Nori uses this in production: All their board decks, sales decks, docs, and even this video are built with HTML and CSS, literally just divs all the way down.

The Breakdown

The world pours 34,000 human years daily into making slide decks, yet agents fail spectacularly at canvas-based tools because they think in language, not pixels. Amol Kapoor argues the solution is simple: stop giving agents human tools and start giving them HTML, the structural medium they already understand.

Was This Useful?

Share