Does anyone know how start up or project like these are built from ground up? I have very very command of react and I can write rest api using spring or express.js. I know finger too about prompt engineering. What does it take to make something like this?
I'm currently building a similar app for my company. Apps like this are built pretty much entirely around prompting. The basic path is:
User Enters Prompt > Combine with Internal Prompt > Feed to AI > Return Results.
That's pretty much it but realize the "internal prompt" will take you months of work and refinement. There are many many tiny edge cases around using AI like this. One problem for example is you can tell the AI to return you it's data in a standard JSON object that you app can use, but if you set the temperature to high then the output will stop conforming. So you start having to do "pass-offs" between conversations with different temperatures.
And this is just one of a litany of issues around using AI like this. And really this is no different than a "traditional" SaaS app that doesn't use AI, I have a side project and that project has a litany of issues that must also be fixed.
Bottom line is building something like this is not any easier than building a traditional SaaS, just different flavors of problems.
I have fun building tiny hobby apps using these AI tools - I built an HN version of Reddit Enhancement Suite where I can tag people, it tracks votes on comments/users, etc, in about 90 minutes using Bolt - but I've tried doing "real" apps and it always seems to collapse under its own weight after the second or third new feature beyond the initial framework.
That has been my experience as well. My top comment talks about building an AI app, but that app is not being vibe coded. I know the article talk about how Base44 was entirely vibe coded but TBH i find that very hard to believe.
The other day I tried to legit vibe code for about an hour and Claude could barely build a weather app that used NWS data for me. Even with this toy weather app it collapsed a few features in, simply put the context windows of these AI tools is just not large enough. I think vibe coding _might_ be viable once context windows start growing into the 5-10M token range but even then I'm skeptical.
