I wanted to try out LocalLLM properly with the news of Qwen3.6 and Gemma4, so
It's been about two months since I bought a MacBook for a lot of money.
I tried various things to make local models do something meaningful.
cline, pi, opencode didn't produce satisfactory results.
It was full of bugs, the UI was a mess, and it took more than 30 minutes to modify...etc.
Yesterday, I made a skill with claude code.
opus plans and breaks down tasks into smaller pieces and writes test code
pi -p "task1 execution" command to execute the task with the Qwen3.6-27B model
opus performs final review
I got the most satisfying result using a local model.
For the first time, a program was created that worked perfectly without any bugs.
But when I asked claude if there was a cost reduction effect, it said there wasn't.
It means that the tokens for planning and task writing are almost the same as the tokens for directly writing code.
In fact, it says I used even more.
I'm a little frustrated about whether this makes sense.
I'm curious how other people are using local models.
