A place for developers to talk about building on Farcaster.

/dev

I don't know how people build things "vibe coding".
I frequently need to stop using the LLM and write everything myself because it keeps messing up.

It definitely saves times but it has also wasted my time quite a bit.

The open web is dead, long live the open web. ⌐◨-◨

Building /warpcast and /farcaster

— gabrielayuso.com

Building https://srcbook.com // creator of the https://blockchainsmokers.xyz

It's great for some things but very hit or miss for anything a bit more complicated.
It often thinks it solved it and it's wrong.

I use Cursor mostly. Either chat or agent mode. Usually sonnet-3.7 but sometimes gpt-o1.

I usually give it the relevant code as context, including the specific lines of code and instructions on what I want to achieve.

It works but not enough to "vibe code" the entire thing especially if it's not basic stuff.

It depends a lot on the task.

Also depends on how you do it.

What tool are you using? Claude code? Cursor agent? With what model? How are you planning vs executing etc

Interested in ZK, FHE, PUFs, hardware, system design, ai safety, and smart contact security.

For new code it's definitely harder to think about the entire flow edge cases if you didn't write it and you're just reviewing it.

I had a lot of experience reviewing code so I'd hope it'd be easier for me but you still miss stuff.

I find it pretty difficult to get a good mental model of what I'm making unless I've written it myself most of the time

I mean the LLM is great. But yea I agree, I have to redirect or just take over at times especially for something intricate. 

Tried making this argument yesterday on FC, and told I was wrong lol

I agree that it's not perfect.

I spent Friday evening forcing myself to not edit the code by hand but rather let Cursor do it. It was rather painful at times, but I also got a prototype that was much more creative than if I had written everything myself.

I think for prototyping it is definitely a tool, but I don't expect it to put human developers out of work in a snap.

It can do simple tasks well, codebase maintenance, docw improvements, simple refactors

I've found it to be best for rapid prototyping more complex algorithms, where you use it first to identify the best type of algorithm to use (since it has an encyclopedic knowledge of data structures and algorithms), and then work with it towards the shape of the solution you want.

Then rip it out and install it yourself lol