Why prompting auth and Stripe from a blank repo still burns evenings
You can ask Cursor or Claude Code to "add Google login and Stripe checkout" on an empty Next.js repo. You will get files. You will not get a product you can charge for that night.
This is not an argument against AI tools. It is an argument against starting the plumbing from zero every time.
What the first prompt actually gives you
A typical first pass looks complete:
- A sign-in page
- A Stripe Checkout button
- A
webhookroute that logscheckout.session.completed
Then you try to use it. Magic links expire in the wrong environment. The OAuth callback hits localhost while Vercel uses a different origin. The webhook secret is missing in production. The customer exists in Stripe and not in your profiles table. Team invites have no role. Billing is not tied to a person who can actually manage the subscription.
Each of those is a follow-up prompt. Each follow-up rewrites a file you already "finished." Evenings disappear in that loop, not in writing the marketing page.
The parts that do not survive a one-shot scaffold
Auth is a system. OAuth, magic link, and password each need callbacks, session cookies, and a rule for who can see /dashboard. If you skip post-signup onboarding, you get users with no profile row and queries that assume one exists.
Stripe is a system. Checkout is the easy screenshot. The durable parts are the customer id, the subscription status, the portal, and the webhook that is idempotent when Stripe retries. Do that math on the server. Do not trust the browser to decide who paid.
Teams are a system. Owner, admin, and member are not a role column you add later if billing already lives on a single user. NextBento keeps the subscription on the owner and lets admins invite people. That split is boring until you get it wrong.
RLS is a system. If the client can write the database, a motivated user will. Privileged work goes through API routes that check the session, then the role.
When a blank repo is still the right call
Use a blank repo if the product is the auth experiment, or you already have a stack you will not leave (Clerk + your own billing, a company starter, a backend that is not Supabase). A kit that fights you is slower than prompts you already know how to steer.
Do not use a blank repo because the agent "should be able to do it." It can generate the shape. You still spend the weekend making the shape survive a real user.
What to clone instead of re-explaining
If you are an indie hacker shipping a Next.js SaaS, start from a repo that already has those systems wired, then let the agent extend your conventions. That is the point of NextBento for AI-assisted founders and the indie hacker page: the villain is re-prompting plumbing, not using Cursor.
The kit is $59 one-time. It is the wrong buy if you need MakerKit-style multi-tenancy or you want a free Wasp stack. Those tradeoffs live on compare, not in this article.
