[albrewer]

My AI experience has varied wildly depending on the problem I'm working on. For web apps in Python, they're fantastic. For hacking on old engineering calculation code written in C/C++, it's an unmitigated disaster and an active hindrance.

[f1shy]

Just last week I asked copilot to make a FastCGI client in C. It gave me 5 times code that did not compile. Afer some massaging I got it to compile, didn’t work. After some changes, works. No I say “i do not want to use libfcgi, just want a simple implementation”. After already one hour wrestling, I realize the whole thing blocks, I want no blocking calls… still half an hour later fighting, I’m slowly getting there. I see the code: a total mess.

I deleted all, wrote from scratch a 350 lines file which wotks.

[paool]

Context engineering > vibe coding.

Front load with instructions, examples, and be specific. How well you write the prompt greatly determines the output.

Also, use Claude code not copilot.

[kentm]

At some point it becomes easier to just write the code. If the solution was 350 lines, then I'm guessing it was far easier for them to just write that rather then tweak instructions, find examples, etc to cajole the AI to writing workable code (that would then need to be reviewed and tweaked if doing it properly).

[f1shy]

Exactly, if I have to write a 340 lines prompt, I could very well start just writing code.

[kortilla]

“Just tell it how to write the code and then it will write the code.”

No wonder the vast majority of AI adoption is failing to produce results.