Using AI to build apps is 1 step forward, 2 steps back.
I give up.
Discussion
They're good for bootstrapping, but production ready applications require understanding each and every aspect of your code
My problem is I'll get some feature working finally after 100 tries, and then it'll do something weird like deleting half of my app views. Constantly doing things I didn't ask it to do or refuses a simple layout change no matter how I rephrase it. Extremely frustrating and not worth it.
Yeah it takes getting used to, one way I get around it is to ask it to write snippets instead of the whole file and then only use those snippets instead of overwriting the whole thing.
What tools are you using to do this? (Or what tools were you using rather)
Lovable.dev
Thank you! Checking it out
This seems pretty slick. Got something basic working quickly, and it was able to add features as I requested them.
What were your troubles with it?
This is why it is essential to (1) modularize everything (2) maintain a mapping and holistic understanding of your codebase. The LLM can't do that, so you need to hold the meta projects and directions in your own head, try to describe it and even convey it back to the LLM itself.
Don't use a hammer to make your concrete 😁
Hold on, I'm cooking something delicious 😋 #osty #ai
Haaaaaalp
🤣
I couldn’t get nostr tools to work beyond a few functions and NDK would always result in some unfixable (by AI) error.
😂😂😂
You don't need to interact directly with nostr-tools or NDK, there is https://osty.dev which will provide ready-to-use full-functional "WIDGETS". you will need to just combine these features together to make things work. hold tight for new releases!
AI generation of anything more than one or two levels deep requires an overarching concept, for the training, testing, and the request formulation.
Or you spend more time fiddling with it, than you would need to just implement a solid framework.
Folks underestimate how much nonverbal communication is involved in software engineering. By the time you can finally accurately articulate what you need, the hard part is already done.
You can push through. You have to understand and use its workflow, provide the most amount of context you think is relevant for a specific output, use multiple chats. The whole point is to carve a channel for its output to flow through. If the channel is too wide (vague prompt will give a wide range of responses) you'll most likely not get what you want. It has no idea what you want, but if you point it in exactly the right direction, it solves things fast. Additionally, it helps you put your feet into the water, getting used to working with whatever tools through automation. Being at least acquainted with it helps you when you want to go straight into the manual to learn.