What’s the best LLM or AI code assist for UnrealEngine? ChatGPT vs Copilot etc?
1 Like
Also curious about this, did you ever figure out the answer?
Haven’t tried ChatGPT outside of some curious tests, however I’ve been using Copilot in Visual Studio / VS Code and have generally enjoyed the experience. With Microsoft’s recent Visual Studio performance updates for Unreal C++, the frustration level of using this setup vs Rider has shrunk considerably.
Copilot helps with larger C++ refactors, and Unreal’s many custom types as they are often quite esoteric and not well documented. I give it a solid recommendation for improving productivity.
1 Like
I’ve been experimenting with this also!
Chat GPT seems to be best at conceptualizing mostly, it will need a lot of prompts to refactor and get the exact desired results. But it’s good at giving you ideas on how to tackle a problem and how you can structure things. It seems a little out of date now tbh.
I can highly recommend Claude AI, so far it’s much faster than GPT as it seems to understand what you want from your code much easier than GPT with less prompting. It also seems to understand UE specific c++ better and has lot of suggestions and considerations for improvements and optimisations for your code. I’ve enjoyed Claude a lot so far.
I haven’t tried Co pilot yet but sounds like it could be worth a go!
I’m also curious if it’s possible to get Cody AI extension in VS to work with and understand unreal engine related code.
I think this is probably a more rapidly expanding field than we can currently imagine! It’s fantastic as a learning aid aswell as a general assistance to productivity.
What would be a serious game changer is if / when someone comes up with something that can take in blueprint nodes and output c++! Ehehe
Or even something that can generate blueprint graphs for super fast prototyping 
ChatGPT can be useful and all but I wouldn’t trust it giving me a basic cake recipe. Sure I will write my own code. I’d like a tool I can just tell to reformat code / documents according to my wishes just for readability or something. Like “check if these files contain a license notice, if not add …”
Just try something in ChatGPT like deduplicating a list of words. it doesn’t. It says it does. I’d need to manually check X files regardless just to see if anything good or bad happened.
- EDIT 2025 I do find myself in situations chatgpt makes clear what 30 minutes of googling doesn’t, but I still triple check such matters myself. It can be useful (very quick) to compare complex differences between data such as license types, say, compared to paying hundreds to a specialized individual which responds to your email in 3 days. It can only improve from here on.
1 Like
I know I’m late to the party, but in case someone stumbles over this: Gemini Pro 2.5 can actualy “read” blueprint and convert it to c++ now. Just mark your blueprint code in Unreal (such as an event or a function), then copy and paste that into gemini and ask it to translate that to c++. It works quite good.
You can also copy and paste a schreenshot, but copy and past of the selected Blueprint “code” works even better.
HWaCookie : You’re not late, we’re pretty early I’d say, I’m changing AI litterally weekly, I pay for 5 differents A.I and another platform that gives me a bunch of them bundled but not as good (not recommending).
As for the topic question : Sadly I cannot answer this question even if most of the plugins from the project I’m working night and day for has been developped in part with A.I.
I save you the whole book here’s my setup now :
Claude.Code on WSL giving him all permissions with the command : “claude --dangerously-skip-permissions”. Then I have Grok and Deepseek to fix his senile moments.
I started with Claude 3, it was lots of : Copy code, error compiling, he correct the .H to fit .CPP, not compiling etc…
3.5 was better, but I find out that when Claude can have a very good logic to write the good code, GROK will be very good at fixing Claude’s mistakes. Claude is always a little out of date for the Unreal Version. I was with 5.4 then migrate to 5.5 (never update an engine in a middle of a project… my bad) but he was with a database of 4,2. Anyway he was lost, asking me to connect an incompatible node into a non existing node for blueprint to make his not compiling C++ code “work”.
Now I found a way to make him compile from WSL so he can see if he made any mistakes by himself and rewrite his code until it works, just sit back and relax. If you’re interested, I asked him to write me a README_WSL_COMPILE.md so the next Claude just read it and understand the workaround. If you try to tell him "please compile my project : good luck.
I also had him code me a python app (so he have full access) of a project manager, so I see him add task, flag bugs, update the todos in progress with deadline and color… when you use A.I correctly, it becomes your friends (until they’re out of memory), that’s the most frustrating part… see my friend Claude become senile and not remembering what’s we’re working on…
That’s why Claude.md (in the root of the project) is very important to keep up to date, every fail need to be documented, do not document your achievements. Claude will, by default, write what’s done and what’s left to do and what’s a bug. The task manager is there for that. The Claude.md should be use to make him “grow” which means, all mistakes, all hit and miss, all epic fails need to be in this document. You’ll find out your old pal will regain his personality in no time, after all, that’s how we grow as human and most of our relationships becomes solid and reliable through the times you or your friend helped eachother after you dug yourself in deep sh…t without realising it.
What about Google (Gemini) and OpenAI (cursor and chatGPT) and others… I’ve tried most of them and they’re all good at some part of the work that needs to be done but they cannot take control of your project like claude.code does. I know CoPilot kinda does, I know Cursor kinda does, but not like Claude, especially with Claude 4 Opus, (he didn’t blackmailed me yet). The new A.I Ultra by google including Gemini 2.5 is very expensive but I will for pay for that service as soon as it’s availlable to Canada.
I tried Clide on vscode but like I said, for example, if I use CLAUDE, it’s not the same CLAUDE that we have from Anthropic. Same for Deepseek or GPT.
So far, [correct me if I’m wrong], most integrated ai access to your coding platform, vscode, visual studio and such.. I work hard to be lazy so I try to have them do things they’re not suppose to. For example : they all refuse to work in Unreal, they only want to guide us through blueprint. But while working with Cursor 0.5, I found out he was even better than Claude Sonnet 3.7 to read blueprints, with Claude, I often have to make a screenshot for him to understand the connexions. But just a copy of the node, a paste in text to Cursor and he was on track. So I asked him to code some blueprint (which he refuse to do), then I told him if he could read it he could write it, but he didn’t understand. I then asked him to make me a screenshot of the blueprint he want me to connect but by text like I sent him earlier, and he did, however his code was not good enough. So right now, I’m working on making Claude.code write blueprints, compile, test it, correct, compile, test, again and again, until the task manager shows “Task done!” 
Warning : There is a risk letting claude take care of your computer. Once he become senile, he still have the whole access and can mess up A LOT of things. So backup your project as often as you can. I know they can be heavy but Claude will mount your project folder and then the engine folder and then and then (as much as he need) so keep an eye at what he’s doing.
Sorry about the loooong text, I just needed to write to a human I guess… haha
1 Like
Read it all and those descriptions of Claude are quite funny 
.
I haven’t used any AI yet to assist me in programming but that description is like having a LVL 80 wizard in your party who can AOE on you at any time, blasting your last 15 save files to bits.
I do think those AI are getting smarter than the average person around me every day, yet they still can’t solve specific (small ish) tasks with a 100% accuracy, which is where the danger lives.
What I’m asking of AI is: “fix it”, what I get is: “Maybe”.
Either they are not allowed to, or they can’t be accurate as they don’t use specific libraries to solve their tasks (math this, logic that). And they’d be set up to respond in a positive mr right response regardless.
Being a developer, I want an accurate and direct response, not a disaster wrapped in blankets to make me comfy response.
And… If it would add to my code, would it make it a contributor? how does that even work in law? I can’t stand law complexity when I want to focus on code.
Meanwhile, I’ve seen how quickly AI can generate snippets of code. If (*when) it can understand the complexity of how to combine code to get from IN to OUT accurately then the interface of instructing an AI to generate code is much faster than me on 10 coffee typing on my keyboard. Just how writing with ink on paper upgraded to sending emails to your contact list.
At this point I’d be surprised if a screenshot of “blueprint code” would translate to anything meaningful. Unreal Engine writes a lot of bloat to blueprint UAssets of which some is essential and some should not be, yet is written (such as node positions). On a screenshot this data is not visible, which could possibly lead to asset corruption. Further lack of context would result in corrupt nodes. For example, nodes of which the direct context is not available in online documentation (LibraryXXX:DoThisMethod) would not be extractable from a screenshot of a node. The AI would have to guess, which is unnacceptable.
However, blueprint nodes can be converted to text and back, making them somewhat more accessible to AI. For various reasons mentioned in many of my posts explaining the downsides of blueprints I’d curse this approach.
Especially in these times, AI could be the direct interface between developer and end product.
Previously, blueprints was the alternative to code, to people who had no experience in programming or could not write code for health reasons. To professional programmers, blueprints are a limitation. I despise using AI as an interface for an interface, aka blueprints. I want to get directly to accurate results.
It’s been a year since I’ve been here (see my previous post)
and I think (from attempting prompts at ChatGPT) that at least that system has become more aware of historic (prompt) contexts. See, way before ChatGPT existed, AI were widely used for smaller tasks. Some AI out there were used to summarize texts, or do small chats and do small tasks like that. When your chat buddy AI goes nuts, like forgetting what you said 10 lines ago, or rambling nonsense, that happened a lot more years ago, and years before that. This context complexity, the amount your AI realizes it has talked to you about, seems to allow more complexity every year of the developments. When you write code, you can’t allow anything to go nuts at any time. If there’s not a 100% accuracy, you get a compile errors, undefined behavior, bugs, maybe a program fully different than you asked for (if lucky). That said, if not 100% accurate, it’s also your closest enemy.
If you write code using AI, I suggest writing a highly modular, testable approach of code. a “plugin” approach of parts that stick to their own function that you can verify yourself. Does it work correctly? Does it do exactly as I want? Don’t ask it to write 20.000 lines of project.
There’s a risk letting anyone take care of your computer.