The past days I had to do several refactoring of codebases. I realized that while AI has sped up the progress of migrating the codes (i.e the bits you rarely like to do), it doesn't take away the other hard bits of figuring out what and how to migrate.
For example, the page of https://lesteuai.com was in SvelteKit 1, I wanted to move to SvelteKit 2 in preparation for using it as the full-stack solution. However, things changed between those two, not to mention its dependencies, with deprecation warnings and wrong syntax everywhere. I let Claude handle the ENTIRE codebase migration. It ran for 15 minutes, running, checking everything, and end up with a page with missing parts and even more errors than I can read. That burned up way too many tokens and still a buggy mess. So with a heavy heart, I decided to scrape the entire result, cause I didn't like what it was given out. The next day, next solution. I moved EACH page at a time, and it went smoother. It still had bugs, but manageable because we fixed as we went. It built perfectly.
Another is a query that ran way too long. As I used Claude to explain what it did, I didn't understand it for many days. Claude was used to improve performance too, but only to a point. When I finally understood, it was much easier to rip apart stuff in the query that was unnecessary.
So, AI is a powerful tool. But it doesn't take away the fact that you are its orchestrator. YOU have to know what you are doing, so AI can know what it is doing. Until people know exactly what they want when prompting AI to build stuff, software engineering is here to stay.