Advice for using AI Code tools - Why AI won't replace programmers

So, I've started using Cursor about 6 Months ago, in the beginning I was quite skeptical but by now I can't imagine going back to coding everything by hand.

Generally it seems to me as though people are still trying to figure out how to utilize LLMs for coding properly, some just flat out refusing to touch anything AI related, others prophesizing how most programmers will be replaced by AI soonish.

So let me start by sharing some experiences I've gathered in these last couple of months.

Be very detailed in your prompts

This is something I've seen quite often, people just write half a sentence and expect the AI to read their minds, sometimes not even writing proper sentences. I've seen prompts like "login page" which the AI generally will completely fail at. Though to my suprise some of these actually did end up working out quite well.

Don't make it change too much at once

This might be me being overly cautious but in general I've found that it's better to change things in smaller steps rather than a big change upfront and then iterating on that big change to make it work (small changes are alright, but things should always generally work!)

Be careful what you wish for

Generally it seems like the LLM is a bit too nice, so far I haven't seen a LLM deny a request, it'll change whatever you ask of it, whether it's a good idea or not.

You must be the architect

Similar to the prior point, you have to make sure to pause every now and then and clean up the codebase, since just like human programmers the LLM also seems to get confused if the codebase grows too much without proper abstractions that ease the mental burden. However the LLM can actually help here, so far it doesn't seem to think up useful abstraction itself but will happily implement ones you decide are beneficial to the overall architecture.

Choose the right stack

At first I tried writing some WebApp with React/Next.js since it seems to be one of the most popular stacks these days, funnily enough the result has been somewhat disappointing. Not entirely sure exactly why but it seems that having to touch so many files for a change confused the LLM. I've gotten much better results when just doing simple SSR with PHP or Express/EJS.

Customize the response style

Now not everyone might like this, however for me it has made things much more fun if hte LLM doesn't just respond like a robot but instead always speaks in rhymes or other stupid things like that. While they won't increase the quality of the code, they make interacting with the LLM more fun and thereby the overall work more enjoyable.

Conclusions

I really enjoy using Cursor, it's made me quite a bit more productive by allowing me to focus on the things that actually matter and let the LLM deal with the boilerplate. There still are many parts that I need to code myself, but that's alright, it does allow me to give the boring work to someone else for the most part. In general it feels like having a super responsive junior dev there to do you bidding.

It also allows me to add some nice finishing touches that otherwise I probably wouldn't bother with because they aren't a priority. Things like nicely written error/log messages. Or add some nice touches to the UI of a prototype.