Explore

How AI app builders like Lovable work

Follow one app from a prompt to paying customers. Each step adds a piece the app builder wires in, and shows who answers for it. About 10 minutes.

Step 01 of 07

A prompt becomes code and a preview

Here's the delivery loop from “How software gets built with AI”, with one outline drawn around the coding agent, the repo, CI and the live app. That's one product: an app builder such as Lovable, Replit, Bolt or v0, a website with a chat on one side and the app being built on the other, running on its own cloud. The developer and the reviewer fade out. In the developer's place is the maker: Priya, operations lead at a freight broker, who doesn't code. She wants a portal for the broker's customers, small shippers. Her chat is the spec. The app builder's coding agent sends it to the model (from a lab such as Anthropic or OpenAI, paid for by the app builder), writes the files of the generated app (screens that run in a browser, plus server code), runs it, reads the errors and fixes them until it runs. The result shows in a preview, a copy of the generated app in the app builder's cloud that only the maker sees, where the loop had a sandbox. Runs it: the app builder. Answers for it: Priya.

Your guess

Priya types: “A portal where our customers sign in, see their shipments and pay invoices.” What's on her screen a couple of minutes later?

Make a guess to continue.

The answer is A. The coding agent writes real code and runs it in the preview, with sample shipments it made up. It looks finished, and everything the prompt didn't say is a guess.

What can go wrong now

The coding agent fills every gap in the prompt with a guess (what a shipment holds, who sees which invoices), and the preview looks finished whatever it guessed.

How teams handle it

Write the prompt like a spec, with completion criteria: who uses it, and what each person may see and do. And keep the plan gate: Lovable and Bolt have a plan mode, so read the plan before it builds. More in AI coding agents: how an agent's task gets done.

Next:The sample shipments live in the code. Real ones need somewhere to be stored: A database, wired in.

Questions people ask

Is the app builder's coding agent the same as a coding agent like Claude Code?

Same idea: an app around a model, with tools to write files, run the app and read the errors, looping until it works. The difference is what sits around it. Here the app builder also holds the code, runs the hosting and usually the backend, and picks the model, from a lab such as Anthropic or OpenAI, whose tokens it pays for.

Can the maker see the code?

Some app builders let her open it, and connecting the project to GitHub puts a copy in a repo the team controls, which syncs both ways (step 5). Seeing it isn't the same as someone reading it, though, and that's the gap step 5 is about.