At work I’m a Software Architect, I still code a little bit but mostly I do architectural work and developers code the things I come up with. I’ve always been writing small tools for myself like a RSS Reader because I couldn’t find one I liked, or a tool which formats JSON when you paste it into it’s window, or a OSX Timer app and many many more. It was always a painful process which took a lot of time and many of the projects I started I had to stop because of lack of knowledge or time.
Some time ago I started using AI to code those tools. About two years ago started just asking ChatGPT to give me parts of the code and was still coding mostly myself, but now I found this tool OpenCode which gives me Agents which can do planning, research, coding and testing after I write down what exactly I need.
They still need a lot of hand holding but I’m learning how to work with them more and more effectively and for example yesterday within a couple of hours I vibe coded Email Forwarder which takes emails from one Email Provider and with help of the IMAP protocol copies them over to a different provider and is useful for synchronizing emails between two providers. My dad has been using this one free provider for a long time but they delete all emails after 6 months, we always struggle to find important older emails, but not anymore because this script synchronizes them with a different provider which doesn’t delete emails.
I also vibe coded this recoder application which recodes videos from my Sony camera and my Samsung and iPhone phones into a format which can be used by Davinci Resolve. That took me quite a long time.
Anyway if you’re curious, Ask Me Anything!
I am a non-programmer but have been doing similar “vibe coding” for some Linux tools as I couldn’t find anything that worked as I wanted.
My question is, do you have any tips/suggestions on a workflow?
I find it really tiring and time consuming when there is a bug/new feature and the LLM’s seem to get stuck in a loop of bad code.
Best advice when you get into a bad code loop is restart by feeding your last working code into a new instance and asking the LLM what this code does. Then ask it if it’s possible to do x. Never ask x and y. Then tell it to output the full new code with x function implemented. You can usually try y after that as well.
One function at a time.
Repeat when you get into a bad code loop.
Have you considered running OpenCode against a locally run model vs cloud-based ones?
I did consider it, I did not test it yet because I only have 12 GB VRAM in my RTX3060 so only small models up to 14b fit in there and from may previous manual tests they were all terrible at coding. I’m using the build in Grok Code Fast 1 model which runs for free online for some reason and it’s reasonably smart.
One of my projects I do though is a Korean -> English translator AI agent which runs local models for the translation planning, translation itself, criticizing the result etc. I didn’t release it yet because the translation quality is still quite poor.
Which do you find it’s easier to work with people or LLMs?
That’s actually a great question, never thought of it. When it comes to simple things then probably LLM, you kind of can convince them to do what you want without them being offended that you’re micromanaging them. But anything bigger they just don’t have the capacity or creativity to think outside of the box and plan long term or anything, so there absolutelly people.
@jeena These are good example projects. But I think if you still do edits and review the code, then technically it’s not quite vibe coding.
I do review code, even had to add a git-commit hook because the agent would not stop committing even if I told it at least 20 times and it was clearly stated as very important at the top of the AGENTS.md But I don’t really do edits, I just tell the AI agent to fix it.
