Don't blame the model for bad graphics. Change the medium: HTML is the agent's native language for visuals.
HTML is All You Need (for Agents to Make Graphics) - Amol Kapoor, Nori · Amol Kapoor
7 min total·Actually worth watching closely: ~5 min·3 must-watch clips
- 0:10 – 1:15Listen
The problem: why agents can't make graphics
Amol Kapoor opens with the core claim: agents are bad at graphics not because of the model, but because of the medium and tools we hand them — all of which are designed around how humans think.
'The model can't do it' is a misdiagnosis; the real cause is a mismatched medium.
The opening is spoken argument with no key visuals — just catch the framing.▶ Jump to 0:10Speaker · Amol Kapoor - 1:15 – 2:10Watch
Tools built for humans are doomed
Breaks down PowerPoint, Figma, and Canva: canvas interaction — click, drag, drop, resize, snap to grid — is built for human hands and human eyes, and the data structure underneath is in a format only the application can read. Existing agent tooling like Figma MCPs, PowerPoint CLIs, and screenshot-and-replace loops still approaches the problem like a human.
Hand human tools to an agent as-is and the output comes out all wrong — overlapping, unreadable, unaligned garbage.
This stretch shows the canvas tools alongside the agent's broken output (visualMoment 89s); watching beats listening to the description.▶ Jump to 1:15Speaker · Amol Kapoor - 2:10 – 3:20Watch
The SVG pelican test: a collective misdiagnosis
Revisits Simon Willison's 'can you draw a pelican riding a bicycle, SVG only' test — the model results really are bad, and the industry (including benchmarks like ArcAGI) reads that as proof models can't reason about space. The counter: asking an AI to hand-write SVG cuts against how it thinks in the first place.
People think graphically; models think in words, tokens, and structure — the test measures a mismatched medium, not a capability ceiling.
The failed pelican examples are slide images (visualMoment 145s); a glance is enough, the argument itself is in the audio.▶ Jump to 2:10Speaker · Amol Kapoor - 3:20 – 4:50Watch
The answer: HTML is the model's native medium
Makes the case for HTML on three counts: tags have meaning built into the language (a heading, a chart, a grid), models have seen and trained on billions of examples, and the browser turns it all into pixels. Rerun the same pelican task in HTML and the result is readable.
The model never places a single coordinate, yet gets charts, fonts, and motion for free — the methodological pivot of the whole talk.
The HTML pelican comparison at 247s is the key visual; you have to see the before and after for it to be convincing.▶ Jump to 3:20Speaker · Amol Kapoor - 4:50 – 6:00Watch
Nori in production: from decks to video
Putting it to work: decouple the editing format from the presentation format — edit in HTML, render to PDF or whatever else as needed. Nori uses it for board decks, sales decks, and on-brand docs, and even this talk video is plain HTML and CSS, divs all the way down.
A slide deck isn't PowerPoint: the audience only sees presentation mode, so pick whatever editing format agents are already good at.
Real production examples appear at 320s, and the 'this video is HTML' self-proof only lands if you actually watch it.▶ Jump to 4:50Speaker · Amol Kapoor - 6:00 – 6:53Listen
Takeaway: think like the model
Closes with the math: something like 34,000 human years a day go into slide decks, most of it fiddling with formatting; give the model access to your company data and it builds the deck end to end, turning 10 hours into about 25 minutes. 'Coding agent' is just bad marketing — give it the right language and it can do almost anything.
Stop thinking like a user. Think like the model — and for graphics, that language is HTML.
The ending is spoken summary and numbers with no separate visual moment; just hold on to the transferable method.▶ Jump to 6:00Speaker · Amol Kapoor