starting to practice a form of vibe coding where I am not allowed to copy and paste or let it modify any of my files directly. I have to manually type everything it generates.
benefits:
1. you understand all of the code, so that its easier to modify in the future.
2. you can make the code your own, fix bugs, tweak it, etc as you are implementing it.
it's a bit slower but the lack of #1 is my biggest complaint about vibe coding.
symbiote programming?
Replies (30)
Probably a much better way to learn and take advantage of this tech
It’s very similar to copying the masters method of learning (assuming the vibe code is something new you haven’t seen yet)
Funny, the lack of #2 is why I type the code out. It's sooo much easier to just do what you want upfront than having to correct the LLM in english after the fact.
Yup. That's where I am mostly at now too. I occasionally do let it do some boiler plate tasks but only with explicit approval on each edit, just because it can do simple stuff way quicker. But in essence, its not trust worthy.
I really like having discussions with it for hours or days on end so we can nail down specifics, edge cases, etc, then just writing the code myself. it's ok for code but organization is where it shines.
definitely
Do you prompt it to give you the code but not add it, or is there a specific coding AI that is easy to do that with?
Code at macro scale, vibe at micro scale.
I don't understand anything when I vibe code. I'm just copying and pasting. It was fun at first but then it got annoying.
If it's just a small website vibe coding is good enough, but if it's a big project I would just hire someone to do it for me. Vibe coding is too much headaches.
I'll probably just zap-work to hire someone in the future. I've hired developers to make me things in the past on freelancer . com website but that website is full of Indians and sketchy developers. Nostr is small so I suspect it's easy to find good developers.
i first put together a development plan, then I get it to implement it bit by bit. after I type it all out, I ask it to create a compact snapshot for the next stage.
I rinse and repeat. i do this all from the chatgpt web ui
although this is a unique project, i'm making it build a wgpu game engine. there's a lot I don't know about building a rendering engine in wgpu so its very informative. learning a lot
Thats what ive been doing for quite some time, it works out really well.
Same
That's so funny. I have the exact opposite experience. I find that it can quickly code well speed components if I give it criteria and feedback, but it totally falls down on architecture.
I've been working as an architect for a long time though, and running teams. So maybe that's why I use it that way: like a team of fast contributors on acid.
*well spec'd
no that makes sense. i will say that it does take some time to understand how it "thinks" and to get it to understand what you're aiming for. that's part of the reason i find it better to flesh ideas and architecture out; it makes stupid mistakes or wants to take weird directions and i rein it in until we're aligned.
Another thing that works well is set permissions to "ask" on each file modification so you can stop and tell it to it differently. After task is done, check all files a and tweak before committing.
I don't think that's what vibe coding means. AI is just a database with a slick interface that does code (and text) generation. If you aren't using it to do things you don't actually understand or are not bothering to check, your not "giving in to the vibes". You're just programming. Which is better, actually.
Yeah, just the back and forth is very useful. I once worked with a guy who kept a goldfish on his desk and talked through everything with it. (Not kidding.)
lol did we work together?
Very interesting thanks
🤣
I wish it was possible for the agents to read/write through the text editor, so I could immediately see the proposed changes and - with a single keypress - tell it that I want to adjust the edit manually
Then the editor should use highlighting to distinguish the changes made by me from the LLM's changes that I haven't (yet) modified
If that's not already possible today (is it?), I guess we'll get there sooner or later
Could you do it in a collaborative coding environment where it inputs the original code and you review and modify it line by line? It could then review and make changes and then you review thd diff?
would be cool, but not sure how useful. it's not hard to just read through it line by line. typing it out just forces me to go slower to digest everything
I don't manually type everything it generates character by character, but I do read it carefully before inserting into my codebase
pair programming
i like this 🙂
Doing manual by hand diffing and copy-pasting the mods makes the code familiar to you but speeds up when you can skip the typing. It takes discipline to actively understand every pasted snippet: a form of meditation.
