A Weekend MVP in Replit AI: What It Built, What I Did
Idea Saturday morning, deployed app Sunday night. An honest ledger of what the Replit agent handled and what a human still has to do.
The idea was small on purpose: a waitlist page with referral tracking and a simple admin view. Perfect Replit agent territory, and a fair test of where the 'weekend MVP' promise actually holds. Saturday morning start, Sunday night, deployed.
What the agent genuinely delivered: the entire scaffold in under an hour. Sign-up flow, database schema, the referral link logic, a clean-enough UI, and deployment with a real URL. That is honestly 70 percent of the work, and it was the 70 percent I would have spent all of Saturday on. The referral tracking, which I expected to hand-write, worked on the first try.
What I did, in order of hours spent: deciding things. The agent builds what you describe, so the description has to exist. I burned real time Saturday afternoon discovering my spec had holes: what happens on duplicate sign-ups, whether admins see emails, what the confirmation says. Every one was a five-minute fix once I noticed and an hour of confusion when I did not.
The agent's honest failure modes: the confirmation email silently failed and it kept proposing the same fix (I eventually read the provider docs myself), and its second-day edits started breaking day-one features, which is the classic agent regression loop. The cure is pinning things: test files, and explicit 'do not modify' comments around working code.
The ledger's bottom line: Sunday night there was a working, deployed, real app, and Monday morning there were real signups. That would not have happened without the agent, and it also would not have happened without a human who spent most of Sunday doing product thinking the agent could not do. The tool is real. The weekend part is real. The 'I don't need to think' part never was.