I Built (And Shipped) a 3D Game With Claude Opus 5.5 (Full Workflow)

Summary Independent developer Chong-U demonstrates how he built and published Pressure Wash Panic!, a fully playable 3D browser and mobile casual game, using Anthropic’s Claude Opus 5.5 and sub-agent orchestration. The game runs directly in the browser via WebAssembly (Rust) and WebGPU without a pre-existing game engine or Three.js. Chong-U details his complete pipeline—from concept art and 3D asset generation to animation rigging, greybox mechanics testing, and final polish—along with cost breakdowns and execution metrics.

What is shown

• 00:00–00:20: Gameplay of Pressure Wash Panic!, showing top-down driveway pressure washing mechanics, surface cleaning percentages, time limits, penalties for spraying flowerbeds/cats/cars, and equipment upgrades.
• 00:21–00:45: High-level pipeline overview: multi-view image-to-3D via Tripo, automated Blender scripting and headless auto-weight rigging, and runtime integration into WebGPU/WASM.
• 00:32–00:42: Social engagement on X (51.2K views) and Wavedash player analytics showing 977 daily active players.
• 03:18–05:15: The "What it took to ship" metrics and cost receipt:
  • First web release reached in 6h 37m wall-clock time, requiring 15 typed prompts, 21 sub-agents, 2.44M output tokens, 1,925 model calls, and an API equivalent of approximately $199.
  • Publishing on Wavedash brought cumulative totals to 7h 55m and ~$233.
  • Full logged spend at time of recording: ~$437 across 25h 24m, 4.43M output tokens, 41 sub-agents, and 3,887 calls.
  • Sub-agent allocation breakdown: Claude Sonnet 5.5 handled background documentation, screenshot capture, and Git commit pipelines, while Claude Opus 5.5 handled core code architecture and simulation logic.
  • External service costs: Fal.ai (GPT Image 2.5 for turnaround sheets and mockups, 32 jobs = $1.94), Tripo 3D (50 credits for character/prop meshes), and ElevenLabs for audio.
• 05:42–06:26: Architecture breakdown showing the 4-layer stack: DOM/CSS UI, TypeScript game flow, 120 Hz Rust/Wasm simulation, and a custom WGSL/TypeScript WebGPU renderer (83 KB binary size).
• 06:27–08:26: Initial prompt and visual exploration generating portrait and landscape mockups across four distinct art styles using GPT Image 2.5 on Fal.ai.
• 08:27–10:50: Parallel agent execution prompt: Agent 1 creates character turnaround sheets in Fal.ai, sends them to Tripo 3D, and auto-rigs in Blender; Agent 2 simultaneously constructs a playable greybox gym to tune water spray mechanics.
• 10:51–11:55: Playable greybox mechanics prototype demonstrating early water jet particle dynamics, surface cleaning decaling, and basic UI controls.
• 12:09–13:16: In-browser model inspection debug tool showing 3D bone skeletons, wand socket attachment, and spring-based aiming physics.
• 13:17–14:16: Documentation generated by Opus 5.5 explaining the "aim rig" physics (under-damped spring mechanics, 120 Hz simulation, inverse ballistics for launch angles).
• 14:18–15:11: Visual polish passes: generating a 360-degree panoramic skybox using image generation, fixing hand-wand mesh alignment, and generating neighboring houses in Blender.
• 15:12–16:35: Implementation of a multi-stage tutorial (First-Time User Experience) introducing fan spray, precision jet spray modes, and persistent oil stain cleaning.
• 17:36–18:11: Outro showcasing an earlier dual-engine port project (Cloudcrest Harbor running in Unity 6 and Unreal Engine 5.8).

Claims & numbers

• The presenter claims the game contains no external game engine and no Three.js, executing rules through an 83 KB Rust-compiled WebAssembly binary rendered via custom WebGPU/WGSL shaders (01:03, 06:21).
• The presenter states the first functional web release took 6 hours and 37 minutes of wall-clock time from the first prompt, using 15 typed prompts, 21 sub-agents, 2.44 million output tokens, and 1,925 model calls, costing an API equivalent of approximately $199 (03:19–03:50).
• The presenter claims the full build up to publication on Wavedash took 7 hours and 55 minutes and cost approximately $233 (04:58).
• The total cumulative spend logged across all iterations was $437 over 25 hours and 24 minutes, involving 41 sub-agents, 94 typed prompts, 4,376 tool calls, and 4.43 million output tokens (05:02–05:15).
• External paid tool costs reported: Fal.ai billed $1.94 for 32 image generation jobs, Tripo 3D used 50 credits, alongside runs in ElevenLabs (04:50).
• The simulation loop runs at a fixed 120 Hz step in Rust/WASM to compute spring physics, hose constraints, and inverse ballistics calculations (13:01).
• The presenter reports achieving 977 daily active players on Wavedash shortly after launching the demo link on X (00:39).

Notable quotes

• "This entire game that you see here was built completely with AI. This includes the game logic, the character models, the environment art, as well as all of the other systems that brought this game to life." [00:20]
• "Stop one-shotting games. They serve a purpose to show capability of the model, but if you're trying to build a game that you're trying to call your own, you definitely do not want to one-shot it." [08:12]
• "Because everything is running in Rust and WebAssembly, this can happen really quickly—it happens at 120 Hz, that's 120 times a second." [13:00]

Assessment This is a detailed, genuine developer walkthrough and technical post-mortem showcasing a playable game built using AI coding and generation tools. The developer presents live gameplay, browser inspector tools, transparent API usage dashboards, exact prompt transcripts, and live repository artifacts rather than simulated mockups or exaggerated claims.

Described by gemini-3.8-flash on 2026-09-29 from the video's audio and frames.