Ask HN: How much of your dev work is vibecoding now?

2 points by sylwester 262 days ago | 9 comments

I am a C, C++ and Go developer and have been avoiding AI for past couple of years. About 2-3 years ago, I have started using Copilot, but wasn't satisfied with the results, so I went back to "classic" development. Things have changed for me around 6 months ago when I have discovered Claude Code. At first, I was using the 20$ version and was surprised how much better it was. I started using it for my personal jobs and clients and then quickly reached the limits and decided to go for the more expensive offers.

Nowadays, I feel like I am mostly just planing out projects in detail with custom prompts, custom MCP servers and just let it do it's work before the code review.

How are you using these tools and how did it change your life/workflows?

As much as I like it, I feel less excited about programming new things.

b_e_n_t_o_n 262 days ago |

Maybe 10% or so. It's nice to quickly prototype something but I end up mostly writing it myself because it never really outputs what I want exactly. And by the time I write tests and interfaces for it, the implementation usually isn't that much more time to do by hand anyways.

I probably need to dive into using MCP servers and such.

haute_cuisine 262 days ago |

I found it useless for code generation as it generates pure technical debt. It works quite nicely for code review as it’s way better at catching boring stuff like typos, type mismatches, and extra opinion is always helpful.

smt88 262 days ago |

I have no idea how you do this. I almost can't get a single good chunk of code from Claude. It use it all day, every day. It gives me unusable code ~95% of the time, and I have to hand-edit it.

sylwester 262 days ago | |

I have C++ and Go projects which have 10k+ lines of code and which have compiled and worked on first try. Are you using a custom CLAUDE.md to instruct it? You can use a global one and also a per-project CLAUDE.md to give it project specific instructions.

Just a day ago, it wrote a ZeroTier userspace network backend for the QEMU/KVM virtualisation platform, which allows to use ZeroTier networks as virtual ethernet devices with proper L2 (e.g. internet wide switch) with VMs - no matter if it is Win95, Linux, QNX, ..

smt88 262 days ago | | |

Which model are you using as your daily driver? I'm on Sonnet 3.7 for now.

attogram 262 days ago |

For work: only AI "on the side" used as interactive docs on whatever issue I'm working on.

For hobby projects: All-in with extensive use of Google Jules to create/improve my stuff!

gooodvibes 262 days ago |

Using AI assistance != vibe coding

sim7c00 262 days ago |

well, i dont use it for code. i tried, really did, but it keeps getting into infinite loops, or crashing itself trying to get out of kernel-level errors by typecasing an error_code to a datastructure it expects -_- (wtf??)

also, last week i had to sit through 8 hours code auditing of 2 parsers... i mean. 2 parsers that should of taken like 20 minutes to review... it was a complete and utter shitshow. functionally worked, for that one test-case -_-. the 2 parsers had a nice 'framework' around it or whatever. a webserver to run them, because why not make a webserver if you can... -_- and each parser as a module in the framework, but they both were architected differently :S and had special handling in the framework for each as a result.. rather than cleaning up the interface for the module -_-.

clearly just LLM trying to do the 1000 monkeys approach to a passed test... i swear i might have been quicker just to re-write the entire thing, but u know, wanted to see if it was even possible to decypher the mess it produced. well, it was. and now i want to live in a hole in the ground somewhere away from civilization.

uncircle 262 days ago |