It's likely you're running into "too deep into mediocre code with unclear interfaces and a lot of hidden assumptions hell" that LLMs are generally poor at handling. If you're running into an inextricable wall then it's better to do a controlled demolition.
ie, take everything written by chatgpt and have the highest-quality model you have summarize what the game does, and break down all the features in depth.
Then, take that document and feed it into claude. It may take a few iterations but the code you get will be much better than your attempt on iterating on the existing code.
Claude will likely zero-shot a better application or, at least, one that it can improve on itself.
If claude still insists on making up new features then install the context7 MCP server and ask it to use context7 when working on your request.
ie, take everything written by chatgpt and have the highest-quality model you have summarize what the game does, and break down all the features in depth.
Then, take that document and feed it into claude. It may take a few iterations but the code you get will be much better than your attempt on iterating on the existing code.
Claude will likely zero-shot a better application or, at least, one that it can improve on itself.
If claude still insists on making up new features then install the context7 MCP server and ask it to use context7 when working on your request.