Comment by Karrot_Kream

I guess it depends? The "refactor" stuff, if your IDE or language server can handle it, then yeah I find the LLM slower for sure. But there are other cases than an LLM helps a lot.

I was writing some URL canonicalization logic yesterday. Because we rolled this out as an MVP, customers put URLs in all sorts of ways and we stored it into the DB. My initial pass at the logic failed on some cases. Luckily URL canonicalization is pretty trivially testable. So I took the most used customers from our DB, send them to Claude and told Claude to come up with the "minimum spanning test cases" that cover this behavior. This took maybe 5-10 sec. I then told Zed's agent mode using Opus to make me a test file and use these test cases to call my function. I audited the test cases and ended up removing some silly ones. I iterated on my logic and that was that. Definitely faster than having to do this myself.