The desktop companion is not the problem that needs to be solved. It seems that every developer who successfully used CC and go on to create some tool for it is doing it for the wrong reasons or because they have to do just something.

CC development is not just development, not all types of development. it’s frontend JS based , and it’s backend development. Only those scenarios work.

Try creating native desktop or mobile app and it’s like a swamp of trial and error.

You have to learn by trial and error what documentation sets and instructions you have to provide at which moment and context and balance at with token cost.. it’s a multi dimensional problem for which there are no recipes that work.

On top of that your direct instructions to not use particular patterns or approaches gets forgotten and ignored y CC with later “you’re right, I should have…”. I am starting to think it’s not solvable by the user by providing docs, examples and instructions. That Claude must have native development baked in to the same level as they baked in the frontend and backend.

What I am getting to is - make a tool to manage those doc sets and contexts and instructions and allow to share those sets between users globally as recipes.

> CC development is not just development, not all types of development. it’s frontend JS based , and it’s backend development. Only those scenarios work.

> Try creating native desktop or mobile app and it’s like a swamp of trial and error.

I had it build an Android app and embedded code for a BLE peripheral at the same time, tailing the logs from each to debug issues. It worked great.

I generally avoid stating in concrete terms what Claude Code “can’t” do. Especially since it keeps getting new built-in features (like the ability to background shell processes and then tail their logs later) that often fix issues people talk about having had a month ago.

It's already a solved problem. Just use GPT-5-Thinking (or Pro) to create a highly detailed spec to execute against. You can also download the entire docs/source code for whatever libraries you're using and tell the AI to reference/search it as needed. Works great.