I've recently started using Claude Code, and also expected to not like it (based on using Claude Web to generate some code). I have 17 years of experience.
But it's actually been fantastic. It's challenging in a new way: splitting things down into small enough descriptions that the LLM don't go down a rabbit hole it can't get out of. While also giving high-level guidance to plan for the whole project.
I like UI design, but implementing it is tedious and slow. With Claude I can get a dashboard page done in less than 1 min, from a spec, that I can iterate on. Instead I can focus on the parts I care about, like backend. I need to nudge it to always care about security and storage concerns. I've seen O(n²) loops where O(n) would be sufficient. It keeps making enough mistakes that I'm not exactly out of a job.
The best way to use Claude Code is to write those descriptions and use examples of what you want the output to look like, so take it in small steps and do manual corrections.
However, you need to have seen some good code before using it to make something good. It's just a tool, like compilers and IDEs. I'm pretty happy I don't have to input programs using eight data switches and a latch switch on an Altair 80. And I'm now pretty happy I don't have to care so much about internal API boundaries, because doing a refactoring is one sentence away.
However, I do keep running out of quota after 2-2.5h out of 5h. If I had a company sponsoring this, it would be a no-brainer to use it more.
This is similar to how I use Claude (bought a Pro sub this year). I like to group up context into the UI and have it either look over my approach or write some skeleton code, then I do the heavy lifting myself.
It's faster for me to do it manually (and I still learn) than having to find hard to spot bugs from the AI output (as it always outputs some weird stuff).