I used to hate vibe coding Every time I tried it, I spent more time fixing AI’s mistakes than just writing the code myself Then something changed We used vibe coding to build a new feature for Viteval (our OSS evals framework) The catch? We spent time upfront making sure there was great markdown context for the AI to reference Here’s what worked: - Added an Agent MD to the repo - Linked in key context (testing, coding rules, architecture) - Let AI take the first pass The result? 90% of the feature was vibe coded. A couple reviews, some testing, and it shipped Do others feel vibe coding is ready for production or still just for prototyping??
Thomas Chaigneau
1mo
Thanks for sharing this on LinkedIn so far I saw a lot of angry posts about vibe coding but I suspect people that post are just victim of their skill issue. I’m having so much fun defining the projects specs and constraints, splitting it into sub-tasks and iterating with Claude-code on it.
Matthew R.
1mo
It’s really good so long as you don’t try to do anything really new. If you need something new, giving it a screenshot really helps. It’s not production ready, but it can allow you to have almost super productivity.
Jacob Davis
1mo
You can get ready for production if you know what you're doing! I'm starting a community for builders and founders who want to leverage it (it's called Beyond Vibes)
Afshan M.
1mo
Also I differentiate between vibe coding and prompt engineering. I find PE more meaningful.
Juan Garcia
1mo
This is how I make it work for me: I break things down into small pieces, simple tasks. For example, “I need a function that takes X input and gives Y output,” or “I need to validate data with these rules.” Sometimes it might be a very simple class I can describe easily. This works fro me because my strength is in design. I can read code easily but I sometimes forget the exact syntax when writing (and this can be faster than checking docs or Stack Overflow). But when it comes to dependencies, especially checking if methods work with current right version... then I know I will have to fix it... almost always. Animations are another pain in the... AI. In the end, what helps me is reducing everything to basic and simple request and committing often, very often so I can roll back when needed.
Ran Mizrahi
1mo
Thanks for sharing! We made this process into a product at Bit Cloud with Hope AI. It gives complete control over architecture, generation of each component and ensures your dev standards resulting at 80-90% accuracy. Give it a try. https://bit.cloud
Oleg Ovanesyan
1mo
I think at this point, AI/LLM can do very focused and well explained things well when it comes to coding. It does not help to think and figure things out, it does help to bootstrap code, automate routine tasks etc. It will improve as we standardize code. There are very little genuinely new scenarios/algorithms, same as original plots for movies. Code is text. Text is only meaningful in context. LLM is great for generating code, but human is still needed for instructions, control and validation. I also think LLM approaches code as it does general text - it repeats itself over and over and uses "metaphors" if you will in code. I think someone will have to create a specialized "software engineer LLM" not for syntax but for engineering especially for enterprise scale and degree of accountability. My opinion may be biased because I experiment with AI to understand what and how it does things. And remember, you can't drag LLM to court, so to speak, so go easy with statutory reporting for SEC, taxes and controllers for heavy machinery:)
Melvin Sommer
1mo
I made a GPT assistant to generate boiler plate for these files: https://chatgpt.com/g/g-686f9b4e8c5481918616911c5f935ec2-codex-assistant You may want to start subdividing documentation into dedicated files. E.g.: - AGENTS.md - General project overview and a cursory description of the other documents - AGENTS.ux.md - Conventions regarding UI design conventions, usage flex grids, what kind of elements should represent different types of information and what properties they need, etc. - AGENTS.orm.md - Full object model with functional-level overview of what each entity represents and what the intention behind them is, lists property definitions, relationships etc. - AGENTS.milestones.md - Near and long term functional results sought with the project, providing a bird's eye context for feature development.
Eden Reich
1mo
With the current SLA of providers I wouldn’t trust any of these yet for production, but it’s getting there. I see some degradation in code generation quality by some major providers. It feels like new users get better responses over users that are using the product for a while - so customers prioritization probably also has something to do with it, not sure exactly. I guess the true reliability comes in if we also add consensus into the mix - where one LLM generates an output and other LLMs are agreeing or disagreeing that statistically the output makes sense - but that alone doesn’t solves yet accuracy and cost also more computing. I guess I’d be happy when the agents are generating 95% the result I was expecting.
Hands-on Technical Leader with Passion & Empathy 🔹 Full-Stack Software Engineer & Builder with Opinions 🔹 Building In Public with Vision & Plan
1moI think its not production ready without the review, control, and fine tuning. But it does provide 20x gains for the boilerplate and, as Matthew R. said, the known tasks. That frees us up to get to the more impactful work in hours and not days. That's why I do architecture-first prompting with all the guidelines and guardrails specified in the first prompt, instead of asking it to dazzle me.