What they did
The poster set one rule for 30 days: no application logic typed by hand. The AI writes; he describes, reads every line, rejects and re-prompts. Config, secrets and dashboard clicks stayed manual, and any time he was stuck for more than an hour counted as a logged break where he wrote the code himself. The product was a small SaaS with authentication, Stripe billing, a dashboard and a public API, scaffolded as a Next.js app with Postgres and Drizzle. The post never names the coding tool, only 'the AI'.
How it went
The product shipped and is in production. Scaffolding and boilerplate were one-shot and the first 40% went faster than anything he had built. He logged nine breaks: duplicate helpers and a second User type because the model could not see the rest of the codebase; a Stripe webhook handler that acknowledged events before persisting them, which would have left paying customers with no record after a database blip; a debugging loop where the model rewrote functions and reintroduced the bug; over-built settings pages and swallowed errors; and the last 10% of edge cases, which took three of the four weeks. He now runs a separate reviewer agent prompted to refute each diff and keeps a human on the merge.
Worth knowing
The author is pseudonymous and the post ends by pitching his own product, xenition, which sells the author, skeptic and human-merge pattern he describes. No repo, product name or metrics are shared, so the account cannot be checked. The Stripe webhook ordering bug is the concrete trap: ask for persistence before acknowledgement and verify it yourself.
GitHub