[hayksaakian]

to back up the author: I'll use ChatGPT with the GPT4 model and it mostly just works by simply asking it to do what I want, the same way I might ask another person.

The main constraints I see right now are around 1) complex projects that don't fit into the token context, and 2) keeping up a "base prompt background info" to describe the general purpose of your program and the technical stack.

Issues I've run into:

1) Sometimes it will do things in a "bad" style and you need to correct it. For example overusing callbacks in code that should use async/await. for JavaScript.

2) Subtle errors around types that the vs code compiler can catch and enable you to fix

3) It doesn't seem to consider execution time, so I've seen it produce wildly inefficient code

[Wowfunhappy]

For issues 1 and 3, I usually have success asking ChatGPT to correct them. Ie "Can you write this with fewer callbacks using async/await?" or "Can you make this code faster to execute?"

[hayksaakian]

Right that's what I mean, it requires "fiddling" to resolve. I think these could both be solved with a larger context, where you could include a style guide.

[Wowfunhappy]

Thanks, I misunderstood.

> I think these could both be solved with a larger context, where you could include a style guide.

Hm, I'm not so sure. If you tell ChatGPT that good code has something wrong with it, it will overcorrect by modifying the working code, usually making it worse or even breaking it!

IMO, this is where you need a human in the loop, to recognize the deficiency and tell ChatGPT what to correct. Maybe in GPT-6 that won't be necessary and human programmers will be obsolete, but not today.