Harness: Am I doing something wrong? Or are my expectations unreasonable

I have tried using qwen 3.6 and 3.8 q5 64k context and I'm just not getting the experience everyone here seems to get. I'm using open code and it will loop, forget what it's doing, and mess up the ui. All I am making is a simple web app that is basically a glorified text turn based game. It shouldn't be that complex.

About a month ago I ended up getting a Claude code subscription. It's been pretty amazing. I don't have any of these problems. I can just give it an iterative improvement and it does it.

Yes I know Claude is a frontier model but I still feel like for something simple like this I should be able to use qwen for his results too. I would much prefer to use a local model anyway. Am I just expecting too much? Should I wait for local models to improve. Do I need a better harness?