When you can’t manage to vibecode it, it means you’re actually on the verge of creating real value
We’ve spent the last two months working on a custom video player for our « cursor for video editing ». Even though it could unlock many new features, it felt unreasonable, for our small team, to spend 2 months on a « non-AI » feature. We tried to vibe code most of it, twice, but it was impossible to get right. We then decided to actually spend the time and while learning about low level video tech, we discovered a whole new field of possibilities. Turns out, not being able to vibe code any of it, was actually a great indicator of where we could bring meaningful value and innovation.
86 views
Replies
Nothing worth having comes easy ;)
Tellers.ai
@sentry_co exactly, to bring real value as engineers, I think we should always search for problems hard to solve with state of the art tools
@rguignar I think there is another flavour of vibe coding that doesnt get much attention atm. Extreme programming. Go as fast as you can on a hard problem. Ignore everything other than getting to a place where there is value for a user. Set a time limit. like. a day or a week. And do whatever it takes to get there. It will be buggy, code will be horrible. But you can fix that later. The point is to find some value that doesnt have parity with anything out there. Then you decide if you want to clean it up and productize it etc. After you get to play with it and show it around.
Tellers.ai
@sentry_co It's actually exactly how we went about it, made a first, buggy and impossible to fix, version and then got a glimpse of what was possible, only then we engineered it properly leveraging our first learnings from the ugly/buggy vibe coded version. Though, for the final version, LLM assistants were almost useless as the LLM kept trying to follow patterns of regular video player implementations.
@rguignar Yeah. Its all what you feed the LLM tho. For more advance stuff, try to lead the LLM. I connect PPLX mcp in my Cline setup. That way I can include more advance research when working on code. Or more low-key. Just do research in PPLX first, then feed the proposal to the LLM. The agent + LLM is at the core just a Logical machine. It is only as good as the navigator, you.
Tellers.ai
@cristian_stoian_urzica Such a nice tool to preserve the flow and avoid leaving the code editor for simple questions. As developers, LLMs shortcomings are great opportunities for real innovations, it's like finding a gold vein while digging through the latent space :)