Currently implemented features:
LLM (GPT, Claude etc) understands blueprints and allows editing, generating, compiling and error handling
Only UE 5.4 supported.
Support for 5.5 and 5.6 by next week.
If you have any features you need, please leave a comment on GitHub.
Did you also created LLM AI that does not confabulate, and just admits it has no clue how to solve problem?
Because that AI trait to always have any reply (no matter correct or not) makes it almost useless as coding help. That and fact it will apologize for mistake then happily do same mistake again in next reply.
No i am not “anti AI amish zealot”, i love AI, however one needs to know limitations of AI.
Yesterday example of why AI is bad outside of training data (blueprints are outside).
Yesterday after 10 years i dusted off Guild Wars 2. I had a question “how do i skin animals for leather working”. LLM AI happily sent me on 3 hours quest to find vendor that sells “Skinning knife”. I could not find it anywhere. Then AI apologized and told me “skinning knife” is quest reward in some really hard to get place. I went there, did that quest, got that knife. And i could not skin animals. So AI made up story (copied stuff from wow into GW2). and explained me how to skin in GW2 based on WOW. Then i got angry and AI said it is sorry, and there is no skinning in GW2, leather just randomly drops from killed animals. 4 hours running around GW2, just because LLMs confabulate when training data is not there. For that LLM it looks like GW2 and WOW are just two separate continents/areas in same game called MMO.
Same goes for coding, and same will go for AI analysis of random blueprints.
Yes, that’s right. LLM is not perfect. I’ve been studying deep learning since 2017 and have been programming for over 15 years. So I know a lot about LLM. But AI coding has come a long way recently. It’s not perfect, but it’s a very good tool that we didn’t have before. And the Claude code using Claude Opus4, which was released a while ago, is also very good.
From what I’ve experienced over the past month, the problem isn’t with the AI coding itself, but with the architecture of the Unreal Engine code. It’s not that the deep learning model can’t learn, but that the Unreal Engine architecture is tangled like spaghetti. It’s confusing for llm. It’s a shame.
Still, thanks to llm, I was able to progress quickly. There were many meaningless and useless efforts, and many situations that could not be solved without programming experience.
This is a separate story from the main text, but it seems like the Unreal Engine architecture needs to change a lot. Blueprints were originally meant to be a convenience for humans, but now they are causing inconvenience.
Thanks for your feedback. Please try my tool too.