Fromprompttoproduct
The stack I use to ship a working AI app in a single afternoon: Vercel, Supabase, Claude.
The three-piece kit
I do not pick stacks for every new project. I have one default, and it has shipped most of what I have built in the last twelve months.
- Vercel for the frontend, the API routes, and the cron.
- Supabase for the database, auth, file storage, and realtime channels.
- Claude for whatever the product actually does that's interesting.
That is the whole list. No queues, no message brokers, no Kubernetes, no Redis. If the project outgrows the kit, I will port the bottleneck. So far nothing has.
Why this combination
Vercel removes deployment from your mind. You push, you get a URL. Preview branches mean every PR is a working environment.
Supabase removes the boring parts of a backend. Auth in two functions, RLS for security, Postgres for the parts that actually matter. The realtime subscriptions are a free tier-2 feature that quietly enables half the products I have shipped.
Claude does the work that justifies the product existing. If you remove it and the app is still useful, you didn't need it.
The flow
npx create-next-appwith the App Router.- Add
@supabase/ssrand configure server + client clients. - Add the Anthropic SDK and one API route at
/api/chat. - Deploy to Vercel. Connect the Supabase project. Done.
The first version of every product I make goes from mkdir to live URL in roughly four hours. Most of that is figuring out what the product is, not building it.
The discipline
The stack is fine. The bottleneck is always the question: what is this app actually for? When that is sharp, four hours is enough. When it is not, neither is four months.
Like this? Get the next post in your inbox — two emails a week, every one shippable.
Six AI coding myths I am done hearing
If you keep saying these out loud at meetups, the senior engineers nearby are quietly rolling their eyes.
