That’s kinda what I mean. Maybe it only works well in some languages, but with the harness I built for C and C++ does a fantastic job of adhering to very strict architecture and style guides. Way cleaner, more readable, better factored, and more interpretable than human generated code, except maybe one or two devs I have worked with. YMMV I guess?

TBF I do burn 200k tokens just preloading the context with onboarding, not including any code, just document trees of development policy documents, style and architectural standards, code and documentation review processes, company ethos and culture, etc. it’s a token fire, but it really works for us.

Also, documentation driven development all the way down.

If you're an enterprise (including startups), you worry about customers, not code quality. There are famously many startups that gained traction despite shit code and then eventually got around to fixing it, to whatever extent was possible, like Facebook HHVM, Stripe's Sorbet, etc.

Ok, and? You can live with that if there are more important things to deal with.

I've stared at ugly LLM code, that I had just had generated, and worked well enough for my purposes. (generally, some quick recursion into a nested python dictionary in order to dig out some property -- especially for linting or quick data analysis).

And I wanted something better, sure, something a bit more readable ...but I just needed it to work well enough to recurse through a yaml file for config file linting, not be battle-hardened against every test case.

So to deal with the mess, I shoved it in a pure function, threw a few basic sanity unit tests around it, put a comment with a disclaimer of "#this is LLM generated code, it is lightly tested, do not use it for anything truly load-bearing without a lot more tests" and I moved on to something else.

Not everything has to be bulletproof.