What It Means to Build a Game with Claude Code in UE5
Building a game with Claude Code in Unreal Engine 5 means using an AI coding assistant to design, script, and refine a fully playable endless runner directly inside Epic’s editor, where the model edits blueprints, issues Python commands, responds to screenshots, and iterates on gameplay logic as a near real-time teammate rather than a basic code generator. In Stefan’s project, Claude Code Unreal Engine work began from a standard third-person template and evolved into a complete endless runner game AI system: a fox avatar sprinting over moving tiles, dodging obstacles, and collecting coins. This experiment goes beyond short code snippets. It shows Unreal Engine 5 AI workflows where the model helps manage tile spawning, lane switching, UI, and collision behavior, turning abstract prompts into game development automation that runs inside a professional tool used by studios.
From Fragile Experiments to a Stable Claude–UE5 Pipeline
Before the playable endless runner existed, Stefan spent more than a month testing different ways to connect Claude Code to Unreal Engine 5, hitting fragile setups that broke or needed constant manual fixes. The breakthrough came from combining two free AI game development tools: UnrealClaude and VibeUE. UnrealClaude gives the model access to the UE5 viewport to capture screenshots and move objects, while VibeUE executes blueprint edits and Python commands inside the editor. Both talk to Claude via the Model Context Protocol, which replaces copy‑paste loops with structured commands. The workflow starts in Unreal Engine 5.7 alongside the Claude Code desktop app, with the plugins fetched from GitHub, libraries installed, and a VibeUE API key unlocking blueprint control. A CLAUDE.md file in the project root records configuration details, and Git tracks every change when the Unreal Engine 5 AI assistant begins modifying core systems.
Designing an Endless Runner: Tiles, Lanes, Obstacles and Coins
Stefan framed the endless runner game AI task with very specific prompts instead of broad instructions. He asked for “a route of tiles that keeps moving ahead of the player and disappearing behind, auto‑forward movement, lane switching with the A and D keys, obstacles that block lanes, and coins that only appear in safe areas.” Claude responded by creating a BP_Runner_Tile blueprint, adding variables for tile length and path length, and tuning spawn timing so new tiles appear as old ones clear out. Lane switching proved harder, because collision checks had to stay in sync with a smooth camera. VibeUE’s Python commands helped adjust these blueprint graphs. The AI also generated UI elements for score, collected coins, and a game‑over screen with a retry button, pushing Claude Code Unreal Engine work beyond core movement into interface and feedback.
Iterating Visually: From Placeholders to a Polished Run
Once the core loop worked, Stefan supplied meshes for a stone bridge, spiky rollers, and collectible coins to replace placeholder shapes. He prompted Claude to swap in the real models, apply proper PBR materials instead of default shaders, and align everything so assets did not clip through railings or the bridge surface. When layouts broke, he took screenshots and sent them to Claude, which used line traces to detect geometry problems and nudge objects into better positions. One pass cleaned up coins that spawned on top of obstacles; another fixed gaps in the bridge. Iteration followed a steady rhythm: open play‑in‑editor, take screenshots, ask the AI to inspect and repair issues such as premature game termination on lane changes or strange speed boosts. Blueprint graphs grew tangled, but the logic held, with human review stopping the graphs from turning into unreadable spaghetti.
Claude Fable 5 and the Future of AI-Assisted Game Workflows
Anthropic’s broader push with Claude Fable 5 shows how far this approach can go. According to The Shortcut, AI researcher Ethan Mollick created three separate playable games from “one initial prompt” in Claude Code, and he wrote that Fable “outperformed basically every other public model I have used by a considerable margin.” These examples, paired with Stefan’s hands‑on Unreal Engine 5 AI experiment, mark a shift from AI as a code toy to AI as a practical game development automation partner. Models that can both build new games and play existing ones hint at rapid prototyping pipelines where designers sketch ideas in text, let AI assemble a working prototype, then refine it inside established engines. The UnrealClaude and VibeUE setup also exposes the real‑world friction points: plugin management, rate limits, messy blueprints, and the ongoing need for human taste and oversight.
