I’m bullish on ai coding assistants, but that doesn’t come without caveats.
The near term is definitely “collaboration” and “assistance”. It doesn’t make a non-developer a developer, but it can cover skill gaps related to narrow domain knowledge vs. broad understanding of why something needs to be done. Think of it like not knowing Spanish, but knowing, in English, everything you need to do to accomplish a task It’s just a chasm to cross. Seems magical, but it’s really not.
Dynamic tutorial creation “show me 3 ways to accomplish X task and explain the pros and cons of each” is pretty amazing.
Most of the currently very focused and neat stuff requires access to good training data. We’ll see how that war gets fought. In other words, old things might be readily available, but less new things. Docs will be there, but will code samples?
Coding assistants sucks at sticking to a version of a package/library/thing if the version you want has a lot less training data. Being outside training dates is a whole other thing. You have to constantly refocus it.
AI, prompted differently, can end up feeling like multiple devs are in your codebase implementing similar things differently. This is very prevalent in UI -> Server interchange where there is almost always 4 ways to accomplish something. Even your context queueing can lead it astray. “I use Laravel, Bootstrap in the frontend and Livewire” each of those have a way to do modal dialogs. Without focused prompting you seem to get recommendations to use each of them at different times.
If you think hierarchically you can get very good focused code and reduce “slop”, but I don’t think slop is an issue. For the most part, if you prompt major elements of code style then you see code you like. You have to test your own code. You can’t get away with not reading it.
In summary, pretend you have a junior contributor who never really learns to anticipate what you want, but is also very very devoted to doing exactly what you ask and is insanely book smart, but sometimes misguided.