From prototype to production: hardening an app built with Claude Code

The best AI-built apps we see are genuinely good products. The gap is rarely the idea or the interface — it is the engineering foundation underneath. "Production-grade" is a concrete, finishable state, and getting there follows a repeatable path.

Step 1 — Read the code and map reality

Before changing anything, we read what the AI actually built: the data model, the auth, the third-party calls, where secrets live, how it deploys. This produces an honest picture and a prioritised list — usually a mix of quick wins and a few things that genuinely need rework.

Step 2 — Foundations first

Data model, authentication, permissions and secrets get fixed before features. These are the changes that are cheap now and expensive later, and everything else depends on them being right.

Step 3 — Make it observable and recoverable

Logging, error tracking, automated backups and a tested restore. The goal is simple: when something goes wrong — and it will — you find out quickly and you can recover. This is what separates "an app" from "a system you can run a business on".

Step 4 — Tests and a repeatable deployment

A modest set of automated tests around the parts that must not break, plus a one-command, reproducible deployment and a clean repository. This is what lets the app change safely instead of becoming untouchable.

Step 5 — Documentation and handover

The app should be runnable and changeable by someone other than its author. Short, real documentation and a clean codebase are worth more than another feature, because they protect everything you have already built.

Where AI fits in this

We use AI heavily through all of this — it is excellent at the repetitive, mechanical parts of hardening. What it does not replace is the judgment about what to harden, which risks matter, and where to stop. That is the human intelligence side of the work, and it is exactly what keeps a hardening project from becoming an open-ended rewrite.