{
  "video": {
    "id": "JRTAtZ5iBkU",
    "title": "HTML is All You Need (for Agents to Make Graphics) - Amol Kapoor, Nori",
    "duration": 420,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "Online"
  },
  "analysis": {
    "video_id": "JRTAtZ5iBkU",
    "title": "HTML is All You Need (for Agents to Make Graphics) - Amol Kapoor, Nori",
    "one_liner": "Amol Kapoor argues that agents aren't bad at graphics — the medium is wrong: give a coding agent HTML instead of a canvas tool and it will produce your slide decks, docs and even videos.",
    "summary": "Kapoor, CEO of Nori, says coding agents are mislabelled — they can do almost anything if you think like the agent rather than like a user. Canvas tools (PowerPoint, Slides, Figma, Canva, and agent bridges like Figma MCPs, PowerPoint CLIs and screenshot-and-replace loops) are built for human hands and eyes, so agent output comes out overlapping, misaligned and unreadable. His counter-claim is that this is a medium problem, not a reasoning problem: a human handed raw SVG would fail the same way, and the pelican test proves it. HTML is the native medium — layout described in structure and tags the model has seen billions of examples of, rendered to pixels by the browser — and Nori builds its board decks, sales decks, docs and videos that way.",
    "key_points": [
      "The world pours roughly 34,000 human years per day into making slide decks, and most of that is fiddling, not thinking — a 10-hour deck should take about 25 minutes once formatting, branding and moving things around are removed.",
      "Every canvas tool (PowerPoint, Slides, Figma, Canva) is built for human hands and eyes — click, drag, snap to grid — with an underlying data structure only the application can read; hand it to an agent and the output overlaps, has no alignment and hides text.",
      "Skeptics cite Arc AGI and Simon Willison's 'draw a pelican riding a bicycle, SVG only' test as proof agents can't reason about space; Kapoor shows the model outputs and calls them 'genuinely deeply really bad'.",
      "His rebuttal: a human asked to handwrite an SVG pelican would fail too — SVG is a wall of numbers, and you can't see from a wall of numbers to a pelican. It's not the model, it's the medium.",
      "Figma MCPs, PowerPoint CLIs and screenshot-and-replace loops all share the same flaw: they approach the problem like a human.",
      "HTML is the right medium — a language incredible at describing layout, trained on billions of examples, with meaning built into tags (heading, chart, grid); the browser turns it into pixels so the model never places a coordinate, and charts, layouts, fonts and motion come for free.",
      "The same pelican prompt in HTML gives a bird in a structure the model can reason about and a human can read, theme and edit line by line.",
      "The editing format is arbitrary — PowerPoint is a tool for making decks, the deck is just presentation mode, and no audience cares how you got there; so pick the format agents are already good at and render to PDF later.",
      "Nori uses the HTML trick for its real board decks, sales decks and docs, and for the talk's own video, which is just HTML and CSS — 'divs all the way down'.",
      "A beautiful deck alone is worth nothing; give the model access to call transcripts and emails and it can build the deck end to end — Nori Sessions, which he's used to build board decks from his phone on the subway."
    ],
    "takeaways": [
      "Stop reaching for canvas tools and canvas-shaped agent bridges (Figma MCP, PowerPoint CLI, screenshot loops) when an agent is producing the artifact — author in HTML and export to PDF or whatever format is needed at the end.",
      "Treat the editing format as arbitrary and choose the one the model is already fluent in; judge a medium by whether the model can reason in it, not by whether humans traditionally use it.",
      "Don't conclude a model can't do something from a failure in a medium a human couldn't handle either — check whether you've handed it a wall of coordinates.",
      "Pair the generation format with real company data (call transcripts, emails) so the agent can produce the content as well as the layout, leaving you the vision and story.",
      "Default to structure and color rather than plain text — plain text is a choice of convenience and usually the wrong one when you're trying to create something of use."
    ],
    "topics": [
      "agents",
      "coding-agents",
      "html",
      "slides",
      "document-generation",
      "spatial-reasoning",
      "benchmarks"
    ],
    "tools": [
      "Nori",
      "Nori Sessions",
      "PowerPoint",
      "Google Slides",
      "Figma",
      "Canva",
      "Figma MCP",
      "SVG",
      "HTML",
      "CSS",
      "Arc AGI",
      "Slack"
    ],
    "quotes": [
      {
        "text": "If you ask me, it's not the model, it's the medium.",
        "at": "02:45",
        "url": "https://www.youtube.com/watch?v=JRTAtZ5iBkU&t=165s"
      },
      {
        "text": "But an AI is not a human. Asking an AI to use a canvas is like asking a human to write SVG by hand.",
        "at": "03:23",
        "url": "https://www.youtube.com/watch?v=JRTAtZ5iBkU&t=203s"
      },
      {
        "text": "What you're watching is just HTML and CSS. It's literally just divs all the way down.",
        "at": "05:31",
        "url": "https://www.youtube.com/watch?v=JRTAtZ5iBkU&t=331s"
      },
      {
        "text": "Stop thinking like a user. Think like the model. Give it the right language, and for graphics, all you need is HTML.",
        "at": "06:45",
        "url": "https://www.youtube.com/watch?v=JRTAtZ5iBkU&t=405s"
      }
    ],
    "words": 1342
  },
  "summary_url": "/#JRTAtZ5iBkU",
  "transcript": {
    "html": "/transcripts/JRTAtZ5iBkU.html",
    "txt": "/transcripts/JRTAtZ5iBkU.txt",
    "vtt": "/transcripts/JRTAtZ5iBkU.vtt"
  }
}