# Building 3D Websites with LLMs: From a One-Line Prompt to an Interactive Scene

    This article explains how large language models are transforming 3D web development, outlines three product tiers for AI-assisted 3D coding, and walks through a complete workflow—using a handmade ceramics studio homepage as a case study—covering prompt expansion, prototype generation, and refinement with reference footage.

    Why 3D Web Is a New Frontier for AI Coding

    The maturation of browser-side graphics standards such as WebGL and WebGPU means 3D scenes now run directly in the browser without plugins. Libraries like Three.js wrap the rendering pipeline, letting developers describe geometry, materials, lights, and cameras declaratively. Even so, traditional 3D web development has a steep learning curve: you need to understand 3D coordinate systems, lighting models, shader principles, as well as animation curves and interaction events.

    The leap in LLM code generation is changing this picture. Take a recently released coding model as an example—its 3D capability is one of its most talked-about highlights. The model can construct code directly, turning shapes, materials, and lighting into a renderable scene, and connect camera movement with page interactions. This means 3D web development is shifting from “hand-built” to “conversational generation,” with developers increasingly acting as requirement definers and quality reviewers.

    Three Product Tiers: Choose Based on Task Difficulty

    Current mainstream AI coding tools typically offer three product tiers suitable for 3D tasks, ordered from lowest to highest operating difficulty:

    1. Browser-based chat: Best for quickly validating ideas. Users describe their needs and the model outputs runnable 3D page code with zero environment setup, but project complexity is limited and output quality is average.
    2. Agent-style work platforms: Suited to medium-scale tasks. These platforms support task decomposition, file management, and multi-round autonomous execution, completing multi-file projects with better quality and controllability.
    3. Professional coding assistants: Best for developers pursuing final delivery quality, where output quality is typically the highest. They support three usage modes:
      • Desktop app: Drive the project through conversation, with a friendly UI—ideal for full-process demos;
      • Command-line mode: Invoke via CLI in the terminal, making scripting and pipeline integration easy;
      • IDE plugin: Embed in the editor, editing as you code—ideal for working with existing codebases.

    The selection principle is straightforward: use browser chat to validate ideas, Agent platforms to build complete features, and professional coding assistants to deliver production-grade code. With each increase in difficulty, the required context description and human involvement grow accordingly, but controllability and final results improve in step.

    Hands-On Workflow: A Ceramics Studio Brand Homepage

    The following case study demonstrates the complete 3D web production workflow using a brand homepage for a handmade ceramics studio.

    Step 1: Expand a One-Line Request into a Detailed Prompt

    A vague request like “make a ceramics studio homepage” rarely produces high-quality results. The right approach is to first ask the model to expand the prompt, clarifying:

    • Visual tone: Warm glaze colors (such as celadon or Ru ware sky-blue), soft diffused lighting, clay textures;
    • 3D centerpiece: A rotating, interactive, draggable ceramic vessel;
    • Page structure: A 3D hero section, a craft story section, a portfolio grid, and a contact block;
    • Interactions: Scroll-triggered camera movements and material feedback on hover.

    The expanded prompt should include a scene description, tech stack conventions (e.g., using a mainstream 3D rendering library), and performance constraints (polygon count, texture size), so it can be referenced repeatedly in later iterations.

    Step 2: Generate a Scrollable 3D Prototype

    Feed the detailed prompt to the coding assistant, and the model will construct the full scene code. The core structure looks roughly like this:

    1. // Scene skeleton
    2. const scene = new THREE.Scene();
    3. scene.background = new THREE.Color(0xf5f0e8); // Off-white background
    4. // Main subject: a rotating vase
    5. const vase = new THREE.Mesh(
    6. new THREE.LatheGeometry(profilePoints, 64),
    7. new THREE.MeshPhysicalMaterial({
    8. color: 0x9fc4bc, // Sky-blue glaze
    9. roughness: 0.25,
    10. clearcoat: 0.6 // Glaze sheen
    11. })
    12. );
    13. // Soft studio lighting
    14. scene.add(new THREE.AmbientLight(0xffffff, 0.6));
    15. const keyLight = new THREE.DirectionalLight(0xfff4e6, 1.2);
    16. // Camera dolly driven by scroll
    17. window.addEventListener('scroll', () => {
    18. camera.position.z = baseZ - scrollRatio * 3;
    19. });

    The prototype stage doesn’t need to be perfect. The focus is verifying that the scene renders, scrolling interactions are smooth, and the page structure is complete.

    Step 3: Refine Details with Reference Footage

    This is the key step for raising quality. Provide the model with screen-recorded footage of the desired effect as a reference, and fix specific defects one by one:

    • Camera: Adjust pacing and avoid abrupt camera jumps when scrolling; use easing functions for smooth transitions;
    • Lighting: Use three-point lighting (key, fill, rim) to simulate a studio setup and highlight the vessel’s silhouette;
    • Crackle details: Simulate the ice-crackle patterns of ceramic glaze using procedural textures or normal maps for added realism;
    • Performance: Limit geometry segments and compress textures to maintain frame rates on mobile devices.

    Key Takeaways

    1. Requirement granularity determines output quality: The gap between a one-line and a three-paragraph request is enormous; prompt expansion is a mandatory step.
    2. Iteration beats getting it right in one shot: Get the prototype running first, then polish the camera, lighting, and materials one item at a time with reference material.
    3. Use tools in tiers: Match project complexity to browser chat, Agent platforms, or professional coding assistants—avoid overkill or underpowered tools.
    4. Performance awareness throughout: 3D web users are on browsers, so polygon count, textures, and draw calls must all be kept in check.

    AI-driven 3D web development is dramatically lowering the barrier to creation. Mastering the “prompt engineering → prototype generation → reference-based iteration” workflow, developers can deliver interactive 3D pages in a very short time that previously took weeks—this is the real efficiency dividend of current technological progress.