One aspect you have to consider is the differences in human beings doing the evaluation. I had a coworker/report who would hand me obvious garbage tier code with glaring issues even in its output, and it would take multiple iterations to address very specific review comments (once, in frustration, I showed a snippet of their output to my nontechnical mom and even my mom wtf’ed and pointed out the problem unprompted); I’m sure all the AI-generated code I painstakingly spec, review and fix is totally amazing to them and need very little human input. Not saying it must be the case here, that was extreme, but it’s a very likely factor.

I just read Steve Yegge's book Vibe Coding, and he says learning to use AI effectively is a skill of its own, and takes about a year of solid work to get good at it. It will sometimes do a good job and other times make a mess, and he has a lot of tips on how to get good results, but also says a lot of it is just experience and getting a good feel for when it's about to go haywire.

One person is rigorously checking to see if Claude is actually following the spec and one person isn’t?

I think it depends on both the complexity and the quality bars set by the engineer.

From my observations, generally AI-generated code is average quality.

Even with average quality it can save you a lot of time on some narrowly specialized tasks that would otherwise take you a lot of research and understanding. For example, you can code some deep DSP thingie (say audio) without understanding much what it does and how.

For simpler things like backend or frontend code that doesn't require any special knowledge other than basic backend or frontend - this is where the bars of quality come into play. Some people will be more than happy with AI generated code, others won't be, depending on their experience, also requirements (speed of shipping vs. quality, which almost always resolves to speed) etc.

It could just be that each of the two reviewers is merely focussing on different sides of the same coin? I use Claude all the time. It saves me a lot of effort that I would have otherwise spent in looking up specific components. The magically autocompleted pieces of boilerplate are a tangible relief. It also catches issues that I missed. But when it is wrong, it can be subtly or embarassingly or spectacularly wrong depending on the situation.

Note that one person is mentioning they use Claude Sonnet, which is less capable than the higher tiers (Opus, etc).

It boils down to scope. I use CC in both very specific one-language systems and broad backend-frontend-db-cache systems. You can guess where the difficulty lies. (Hint: its the stuff with at least 3 distinct languages)