Making an AI assistant check its own work in Unreal Engine 5

UE-cortex
I build Unreal Engine plugins, and I wanted to see how far an AI assistant can get on a new UE5 project if it works with a plan and gets checked, rather than just chatting. The result is ue-cortex, a plugin sold on Fab. This post is about the method, and about where it does not work.
ue-cortex does not call a model. You bring your own assistant (Claude Code, Cursor or Codex). The plugin gives it 11 specialist playbooks (systems C++/GAS, multiplayer, world building, Blueprint, DevOps/build, UI, audio, cinematics, performance, technical art, AI) and 83 reference files, and it checks what happens in the editor.

1. Game design document first. The run starts with cortex_start and the assistant drafts a GDD. The plugin validates it: every feature needs testable acceptance criteria and assigned agents. For a small example, "Use ue-cortex to build a dash ability with a stamina cost" becomes a GDD with acceptance lines like "Shift moves the player 6 m", "costs 25 stamina" and "1.5 s cooldown". The GDD is written into the project as Docs/UECortex/GDD.json and GDD.md, so you can read and edit it before anything is built.
2. An approved plan, split by specialist. After you approve, cortex_approve_plan builds a dependency-ordered task graph, one step per agent the GDD uses.

What it does not do:

• It covers roughly half to two-thirds (50-65%) of the engineering scaffolding of a new project.
• Art, level design, tuning and polish are human work.
• The assistant cannot play the game and cannot create a brand-new level.
• A generated project is a compiling, running starting framework, not a finished game.
• It is tested and ready on UE 5.8 only. UE 5.6-5.7 is a possible later version, not a commitment.
• Compile verification is Windows-only, because it uses Live Coding.