Tooling: started developing in Replit including hosting, added Cursor later when it got a bit more complicated.
FE: React SPA with Tailwind; I don't think he was knowledgeable enough to pick a UI component library.
BE: Database is Supabase, auth is Firebase. He had a hard time getting the AI to work with Supabase auth so I suggested Firebase instead (Firebase auth is dead simple and lots of docs). This created another challenge which was using the claims from Firebase in Supabase for the row level security. Rest of the backend is running in Replit autoscale deployments. Stripe integration for payments and other integrations for common third party systems in this particular industry. All done via AI and handing it relevant documentation and code samples.
Replit's tooling is interesting because it can run the app and see the errors in the console. It can also take its own screenshots to see if it's getting it right. I watched some of his sessions and the LLM often makes mistakes. He'll try a few passes to get the LLM to fix the mistake, but if it doesn't work, he moves and will come back to it if it is non-breaking.
I talked to him a bit more and he said he thinks his "super power" is managing an offshore team in India for a few years so he's very used to managing features and code quality via written instructions. Where an engineer (like me) might just give up and take it into our own hands when things go wrong, he can just keep chatting with the AI and pointing it to docs and such.
