ChatGPT is almost useful for whole-picture coding

Frickin' ChatGPT (even ChatGPT 5) is so damn frustrating.  It's almost useful, to the point where it's useless, because it chews up so much time while you try to push it from nearly right to actually right.

I've enlisted its help in doing some website coding, which is a combination of Django on the server, and htmx in the browser.  (If you haven't met htmx, it's an awesome concept)
I wouldn't need ChatGPT, except that while I can code Django in my sleep, I'm terrible at front end coding.

So, I carefully specify the requirements, and ChatGPT happily emits 15+ files it says will do the job.  And they look great.

So I spend the time to copy the contents of all those files to VSCode, and run it.

Soon, bugs emerge.  I ask ChatGPT about the bug, and ChatGPT says "yep, sorry about that, here's a new file".  Which then has another bug in it, or it's inconsistent with some other file it already gave me.  So, I ask ChatGPT to generate me a new set, which it does.  But it's forgotten what it did last time, and there are differences between how it structured things last time compared to this time.  So I have no choice but to copy across the contents of all 15+ files.

Rinse and repeat a few times, and then lo, ChatGPT starts forgetting requirements I already specified.  When asked about it, ChatGPT says it assumed I didn't need those requirements any more.

I'm fed up with however many cycles of stuff that almost works, which then ends up with me wasting 20 minutes finding out how it doesn't actually work this time. 

I'm going to mention here, that a concept from Stalinist USSR was the concept of "wrecking": Something somewhat different to sabotage, it means intentionally or unintentionally doing ones job in a crappy way so that the organisational objective couldn't be met.  A kind of "silent quitting" passive aggressive fouling up the smooth operations of things.  

Wikipedia: Wrecking (Soviet Union)

Stalin asserted that a good deal of the failure of Soviet-style socialism to achieve its goals was because of the presence in the work force of "wreckers", people who were against class struggle.  A kind of economic fifth column, although not necessarily organised.  

I'm going to suggest that LLMs are a form of wrecking agent, designed to prevent those who would trust it from getting any successful work done by emitting ostensibly useful but actually flawed information.

Who would be behind such a thing?  I'm not sure there was any particular evil overlord behind Soviet wrecking; it was just a convenient bogeyman on which to blame failure to meet utopian standards.

Or perhaps communists, or aliens, are behind it.  Communist aliens even!  I don't know.  All I know is I still don't have completely working code that meets my requirements, and I'm pretty sick of wasting time trying to get ChatGPT to provide it.

Update: I've recently been using DeepSeek.  It's miles better.  It's rarely wrong, and it does a much better job of understanding the whole of the problem, as if its attention span was better.  DeepSeek is actually useful.  Oh, and the chat mode is free, so you can talk and talk, and it won't suddenly say "well that's it for today, unless you pay me money".