I read an assertion by somebody that they told Claude what to do and a half hour later it had done the job. I saw an actual PhD person say that she has AI do programming for her. Some youtube expert named Nate talks about how Claude is doing tons of work for him. News reports say that hiring is off because of AI. I just do not see it.

I note that I am not a newb on any level. I have been programming for decades and know a TON about AI and am very thoughtful about using it. I am not doing anything stupid so, please, don't respond as if this is a 'me thing'.

One example happened last week. I spent literally a couple of hours working out a good design and implementation plan for an addition to a very complex CLI tool I have made. My premise is that I want to have documentation as good as I would insist on from a young programmer when I was supervising young programmers so that I could be confident they understand the problem and how to implement.

I had Claude write the docs to files, cleared context and set it to work. It cranked along, with some checkpoints, off and on for most of an hour and, Voila!, it worked.

It turns out that this was actually step one. Having gotten it work with the existing underlying tech stack, I wanted to refactor to use new tools that I have developed separately and that are in production working. I did another long planning session, got docs written, cleared context and set Claude to work.

After awhile I see that Claude is having trouble. It's testing and not getting success. Adding log statements. Fretting, reconsidering. It's not able to figure out why a value isn't available where it needs to be. I see it thrashing and stop the process. Then I look at the code for the first time in this session.

(Remember my reading about people setting Claude to work and getting results. This was as least partly an experiment in that.)

It was absolutely incomprehensible. It had a thing where it defined a function, passed it to a 'control' module and then brought it back unchanged for execution. Many of its variable names were completely unsearchable things like 'value'. It wrapped everything in try/catch blocks but often had no logging to identify what failed. The list of bad examples is endless. It was such a mess that even Claude could not figure out what it had done.

Eventually, I was realized that Claude had ruined my program. Still, it worked. So, I gave up on step two (reverted the repo to step one) and was sad. I spent the next full day refactoring like crazy to make the code something that I could understand. A full day!!

Now, this is not to say that I do not get value from Claude. A little while ago I said, "There is a function on line 127 of file X that I want converted into a separate module." I told it a few details and it worked great. I had a web UI thing where I said, in the listing component of page Y, I want you to make the column headings 'click to sort' and, it worked great.

But I am starting to think I cannot have Claude to substantial primary coding.

What are your experiences?