I’m working on a reciprocator-style spray simulation in Unity and have hit a wall using ChatGPT-4o. The tool fails consistently on the following:
Can’t interpret UI screenshots
Breaks down during Unity scene setup and animation logic
Repeats mistakes or loses context
Hallucinates Unity menus, features, or script methods
Doesn’t hold up on anything beyond basic code snippets
I’ve spent way too long correcting false guidance and trying to work around its limitations. I’m done with GPT for this type of work.
What I’m looking for:
A conversational tool or assistant that can work inside the Unity pipeline
Real support for animation triggers, MonoBehaviours, simulation logic
Consistent scene guidance and Unity Editor familiarity
Something that gets better over time, not worse
Tools I’m evaluating:
Claude 3 Opus – good with logic, but doesn’t execute code
Phind Pro – strong for finding real-world examples
GitHub Copilot – solid for C# code completion, but not much more
Replit/Ghostwriter – not usable for Unity
Crew AI – not impressed so far
Question:
Has anyone here found a tool that actually helps with Unity development beyond basic game scripting?
I’m looking for something I can rely on for simulation work, not another loop of vague suggestions. Would love to hear what’s actually working for Unity devs building more than basic click-to-jump games.
There are many tools that can assist with development, and this isn’t limited to AI tools. However, they are just that, tools. They help developers who already know how to do something accomplish it more easily or efficiently. No tool can replace a programmer’s knowledge or write code for tasks the programmer doesn’t understand how to do themselves.
What truly works is knowing how to code something yourself without relying on any tools. Once you have that foundation, you can integrate tools into your workflow to boost productivity, speed up iteration, or offer a fresh perspective, especially when you’re stuck viewing a problem from only one angle.
Being able to do things without AI assistance helps you avoid many common issues, because you’ll be able to recognize and fix problems as they arise. No tool can replace missing knowledge or reliably produce correct code for something you don’t know to do by yourself.
You can’t rely on any tool, only on yourself.
Focus on learning how to do the work yourself first. Once you’ve built that foundation, use tools to enhance your efficiency. The tool’s limitations won’t be an issue if you already know how to code the solutions, just use the tool where it performs well.
There’s currently no tool that can fully replace a programmer. All tools have limitations and should be used to boost productivity in areas where they’re effective, not relied on in areas where they fall short. Handle the tasks that tools like ChatGPT can’t do on your own, because every tool will have similar issues. Whether it’s in different areas or the same ones, you’ll always encounter false guidance, hallucinations, and other problems.
All LLM based AI tools hallucinate, you won’t find any that are able to replace a developer. And no LLM can actually simulate Unity engine.
Also, OpenAI 4o is not the right model for the task. o3 has the best results in my experience, you want the reasonings models that do multiple step post processing of the prompt.
google gemini works fine for me for godot, it should know unity and c# also. but you need to know what you are asking and be specific in the prompt.
asking the AI: “build me a rpg in unity” you will get subpar results. kindergarten answers.
but asking the AI to focus on specific parts of the game with clear guidelines to follow that will give surprisingly good results especially for difficult problems. I’ve not tried but I’m assuming the paid apps mentioned above do exactly this, rewrite the prompt in the back end to help you focus on specific parts, which probably can help if you really have no idea what you are doing.
JetBrains AI doesn’t rewrite the prompt. At least not likely in the way you’re thinking. It does what many AI tools intended for IDEs do: optionally try to supply context to the code you’re asking for assistance with. For example if I have a class that I need help with that is referencing other classes in my project it will try to include those classes in the prompt. It may even go deeper and include ones that they reference if they’re somehow relevant to the code I want help with. It’s generally very successful at this as its ability to identify code like that was in there since before LLMs became mainstream. It’s considered to be one of the best for a reason.
You do have to enable it for a prompt though (either by typing “@codebase” or clicking the appropriate button) as it does eat into the token quota. Granted that quota is absolutely enormous. I’ve barely touched mine even with hundreds of prompts.
