Unreal Engine and AI Coding: My Hilarious Coding Misadventures

The recent push toward AI video generation felt like a massive shortcut, but I realized I was just playing creative roulette. You hit a button, wait for a result, and hope it matches the vision in your head. When it doesn't, you realize you have zero levers to pull to fix it. That lack of control is exactly why I’ve moved back into the heavy lifting of Unreal Engine 5.7. In an engine, if something is broken, you can actually reach in and fix the logic.

This update covers the messy reality of trying to merge traditional game design with modern AI tools. I’m currently wrestling with "vibe coding" using Claude Code and GeminiCLI to handle the friction of technical implementation. The goal isn't just to make a game; it's to build a world where NPCs aren't just reading from a script. By integrating local LLMs, I’m trying to create characters that can actually hold a conversation, even if they currently spend half their time walking into walls. It’s a transition from being a "generator" back to being a "creator."

Key Discussion Points

• The Illusion of Speed: Why AI video tools offer a fast start but often lead to a dead end because of the "black box" nature of the generation.
• Returning to Unreal Engine 5.7: Reclaiming technical control and why the complexity of a real game engine is a feature, not a bug.
• The "Vibe Coding" Workflow: Using Claude Code and GeminiCLI as high-level pair programmers to bridge the gap between a design idea and working C++ or Blueprint code.
• Local LLM Integration: The technical hurdles of hosting large language models locally to power dynamic, non-linear NPC dialogue systems.
• Solving the NPC Glitches: A look at the "coding shenanigans" involved when your AI characters have great things to say but still can’t figure out basic pathfinding.
• Version Control as Sanity: Why Git is the most important tool in the stack when you’re constantly breaking your project with experimental AI scripts.

Technical Stack & Ecosystem

• Engines: Unreal Engine 5.7 (Primary), Godot (Contextual comparison).
• AI Tooling: Claude Code, GeminiCLI, Local LLM frameworks.
• Concepts: Procedural Generation vs. AI Output, Human-in-the-loop Design, NPC Behavioral Logic.

• Contact for collaborations or feedback: [e-mail]