{
  "video": {
    "id": "VBCDhRrvlYo",
    "title": "The Next Game Engine Won't Have a Manual — Arturo Nunez, Nereu",
    "duration": 1173,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "VBCDhRrvlYo",
    "title": "The Next Game Engine Won't Have a Manual — Arturo Nunez, Nereu",
    "one_liner": "A 10-year Unity veteran demos Nereu, a browser game engine with no scripting layer where an LLM assistant edits games by adding and removing intent tags (an \"Asset Tag System\") instead of writing engine code — arguing the context should sit on game design, not on the engine.",
    "summary": "Arturo Nunez argues that bolting LLMs onto Unreal or Unity just builds a bad bridge between two worlds: you still need engine and code vocabulary, and the model reinvents the wheel every time you ask for something as standard as a follow camera. His alternative is to flip the context from the engine to game design — everything is an asset, you attach tags that describe intent (\"character\", \"animated\", \"double jump\", \"drivable\"), and built-in systems query those tags, an idea lifted from ECS / data-oriented design and called ATS, the Asset Tag System. He demos building a playable scene in minutes — robot character, WASD movement plus animation, buildings, rain as a configured particle system, follow camera — purely by describing it to an assistant, and closes on how he assembles context using level-of-detail, the same trick games use for rendering.",
    "key_points": [
      "Live demo builds a game from plain description in a couple of minutes: find a robot asset, make it move with WASD and animate it, drop in buildings (warehouse, store, castle), spawn rain as a pre-configured particle system, and add a camera that follows the character — no coding and no asset importing.",
      "The complaint about today's LLM-plus-engine setups: \"by default, the context is on the game engine rather than on the game design part\" — ask for a camera that follows a character and the LLM reinvents the wheel every single time, even though the result is always essentially the same.",
      "In a conventional engine a controllable character needs a mesh, renderer, animator, rigid body, collider, audio source, then movement logic and game rules — mostly boilerplate every game already has, yet developers must read component docs and hundreds of sliders.",
      "ATS (Asset Tag System) derives from entity component system / data-oriented design: assets carry tags, systems query for them (\"move everything with the vehicle, player and drivable tags\"). A building has no tags so it doesn't move — add vehicle and drivable to it and you have a drivable building for a Mario Kart-style game.",
      "The agent loop builds its prompt from scene context plus extra context on the asset style in use (robots, medieval) and the user's own description of the game they want; the agent then calls tools that append or remove tags. All tags and systems are built into the engine.",
      "There is deliberately no scripting system — but it's JavaScript running in the browser, so anyone who wants to extend it can; for most users that shouldn't be the case.",
      "Context assembly uses level of detail, borrowed from rendering: a scene had ~100 assets, most of it ignorable grass. Assets near the one being edited (the knight) get full tags and setting values; nearby ones get only \"there's a player here at this position\"; it re-updates as the user moves around.",
      "The 6,000–7,000 assets were too many to tag by hand, so he screenshotted each 3D file and ran a vision model to describe it; the rest of the system is mostly a plain LLM.",
      "He built it with Claude Code, bringing in domain answers from industry friends (e.g. lighting), feeding each new tool's definition to the assistant as it's added, and actively deleting settings and features he hasn't seen users use. It's in closed alpha with a waitlist.",
      "The hard part isn't the AI — it's decomposing genres, and things like a game's mood (\"a platformer that feels scary\" touches post-processing and lighting), into tags and systems.",
      "On world models as an alternative path: he thinks it'll be a different medium even if we call them video games — 60fps at 4K in real time plus simulated physics is still far away."
    ],
    "takeaways": [
      "Put the context on the domain the user actually speaks, not on your engine's internals. Tags describing intent — \"character\", \"animated\", \"double jump\", \"when you collect a coin increase the score\" — are the vocabulary players already learned from tutorials.",
      "Make recurring results into built-in systems rather than letting the LLM regenerate them. If a follow camera always ends the same way, expose it as a tag the agent applies; don't let it write the code again each session.",
      "Apply level-of-detail to context windows: send full tag values only for assets near what the user is editing, positions-only for the periphery, and drop the noise (the grass) entirely — then refresh as focus moves.",
      "Use vision models over screenshots to bulk-tag an asset library you only have filenames and 3D files for; it made 6,000–7,000 assets searchable by description.",
      "When shipping new agent tools, feed the tool definitions to the assistant as you add them, and prune settings and features you don't see users touching — simplification is part of the build loop."
    ],
    "topics": [
      "agents",
      "game-dev",
      "context-engineering",
      "tool-use",
      "vision-models",
      "no-code",
      "world-models",
      "ecs"
    ],
    "tools": [
      "Nereu (Nereo)",
      "Bibi / Weeby (the in-engine assistant)",
      "Unity",
      "Unreal",
      "MongoDB",
      "Claude Code",
      "JavaScript",
      "vision models",
      "LLMs"
    ],
    "quotes": [
      {
        "text": "By default, the context is on the game engine rather than on the game design part, and I think we should flip that idea.",
        "at": "07:41",
        "url": "https://www.youtube.com/watch?v=VBCDhRrvlYo&t=461s"
      },
      {
        "text": "I've seen the demos, and the LLM reinvents the wheel every single time. Well, when the result is going to be essentially the same.",
        "at": "07:35",
        "url": "https://www.youtube.com/watch?v=VBCDhRrvlYo&t=455s"
      },
      {
        "text": "Most of that is just boilerplate that every single game out there has or every character in every single game out there has, but somehow developers need to understand and read the descriptions of those components and what the hundreds of sliders do.",
        "at": "08:09",
        "url": "https://www.youtube.com/watch?v=VBCDhRrvlYo&t=489s"
      },
      {
        "text": "I don't want us to one-shot games that nobody is going to play and I don't see the point in that.",
        "at": "16:49",
        "url": "https://www.youtube.com/watch?v=VBCDhRrvlYo&t=1009s"
      }
    ],
    "words": 3668
  },
  "summary_url": "/#VBCDhRrvlYo",
  "transcript": {
    "html": "/transcripts/VBCDhRrvlYo.html",
    "txt": "/transcripts/VBCDhRrvlYo.txt",
    "vtt": "/transcripts/VBCDhRrvlYo.vtt"
  }
}