Build a small SaaS with Lovable, Supabase and Stripe using a living spec file and a habit of checking the database yourself

AI app builders are fast, but they can ship silent failures such as a checkout that charges without recording the payment.

Keep one spec file covering features, business rules, constraints and past failures, and have the model read it before each build. Build one feature at a time and update the spec in the same session after shipping. After each feature, check the database rows and logs directly instead of trusting the interface, and audit billing edge cases, row-level security and edge function logs. Use the product's own data to decide what to build next, for example topic mismatches in collaboration requests and repeated requests for features that already existed. Do not hand off high-stakes systems such as large-scale money movement or regulated data.

She reports shipping a topic-matching function within a week, a draft generator after approved workspaces showed zero edits, and a release-notes pipeline in days after finding about 40 percent of feature requests were for features that already existed. Audits caught a checkout that charged without writing to the database and an upload that reported success but saved nothing.

Self-reported in an interview with no data shown, and the founder promotes her own products. She says the builder is fast but prone to silent failure, that AI tools report success without verifying it, and that AI readily endorses whatever feature it is asked about, so it does not improve decisions. Substantial time goes into verification.

Accounts with Lovable, Supabase and Stripe, and enough technical comfort to read database rows and logs.

• AI Workflow Stories on Substack (interview with Elena Calvillo) ↗ Firsthand write-up, published August 13, 2026

Source published: August 13, 2026. Last reviewed here: October 13, 2026. Spot a mistake? Tell us.