Vibe Coding Help

[apoyusiken]

So I had been vibe coding mostly by solely getting llms to write the code (it wasnt easy to modify and let llm know of it) and it was harder than good old coding (but it saved me of doing uis which i despise). Recently i ditched that and only use llm for help and this has been incredibly wholesome. Now the question is, what tools would you recommend that lets llm look at the project and suggest stuff etc. besides gpt canvas? 

[whispous]

None. Write it so you know how it works, so you can fix it.

[Timme]

The thing with vibe coding is that when used out of context, your prompt basically needs to be a design doc for the thing you want in code. Some say that the vibe coding is at it's best as a context-aware helper with a big project and automation of trivial coding chores.

[jaslion]

As someone who keeps seeing vibe coders get fired. None! Those people are worse than no help at all in projects and in general cause more damagae that needs to be fixed later.

 

If you write your own shit code and document it then at least you can go back later and go damn I wrote GARBAGE lemme fix it. Instead of the junk machine making a snippit and you having no fuckin clue if its gonna be a problem or anything

[apoyusiken]

ok guys appreciate your comments but i guess i was not clear enough. I know vibe coding sucks as of today, but imo not using ai is as bad as solely using ai..

[colonel_mortis]

I have had good results on small projects with Claude Code (and reasonable results on large projects, when I'm very careful with the prompting). I have also heard good things about, although have not personally tried, OpenAI's Codex. If your previous setup was just to iterate with a regular chatbot LLM, I think you'll find any of the agentic coding tools a big step up - they are much better at editing and take longer to get confused and stuck.

 

Pure vibe coding, where you get the model to just write a whole piece of software, is fine for proving a point or building simple things, but is not good for building long-term maintainable code. However, LLM-assisted software development, where you take full ownership of the code but have the LLM automate some of the actual writing of it, or use the LLM to experiment, can in my opinion be an effective way to write code a bit more quickly, without sacrificing much. For example, at work I sometimes use it for things like writing tests, helping with refactoring, or trying out an idea, but the actual production code will be mostly or entirely written by me. On some personal projects, which are much smaller, I have had LLMs do all the work, and me just supervising, with decent success.

[Nayr438]

If you mean like code review and templating, I'm going to recommend Cursor with Anthropic Claude, if you want it to be actually usable you pretty much need to be on a paid tier.

Just be aware that while AI can produce functional code and can definitely increase productivity, it often gets things wrong and over complicates things. Just because it works it doesn't mean its correct and just because it thinks something is better it doesn't mean it actually is, it's up to you to figure out if it's right or wrong.

Ive had these models create entire c++ classes for something that could be done in a few lines of code using the std library in the most inefficient way because it thought it was better, at the same time Ive seen it do the opposite as I can often over complicate things as well. You need to understand what it's doing and how it impacts your project. Pay close attention to everything it does as it can delete and rework areas that you never wanted it to touch, may even try to entirely add or swap out libraries.

[apoyusiken]

4 hours ago, Nayr438 said:

If you mean like code review and templating, I'm going to recommend Cursor with Anthropic Claude, if you want it to be actually usable you pretty much need to be on a paid tier.

Just be aware that while AI can produce functional code and can definitely increase productivity, it often gets things wrong and over complicates things. Just because it works it doesn't mean its correct and just because it thinks something is better it doesn't mean it actually is, it's up to you to figure out if it's right or wrong.

Ive had these models create entire c++ classes for something that could be done in a few lines of code using the std library in the most inefficient way because it thought it was better, at the same time Ive seen it do the opposite as I can often over complicate things as well. You need to understand what it's doing and how it impacts your project. Pay close attention to everything it does as it can delete and rework areas that you never wanted it to touch, may even try to entirely add or swap out libraries.

yea mfs be like i got llms to write 2k lines in a few hours, bro it wouldnt be 2k lines to begin with if it wasnt for llms..

[Franck]

Github Copilot with Visual Studio you can tell him what to look at specifically. You can target solutions, project, files, methods, class, selection as a source for the prompt. Very useful to be able to pick what you want. In my case i don't use for vibe at all, i use it for the convenience tools it brings. On of my most used part of it is the better auto complete intellisense. I know want i want to write and when making large changes it detect the pattern i am fixing and replicate my previous manual fix when i bring the cursor to my next target where i want to apply similar changes.