Lovable generates fast, but editing is another story

Let me say something. Last week I had a second-year master's student try out Lovable to build a small internal tool for our lab, managing image annotation task assignments. He started from a one-sentence prompt and in just a few minutes got a working page, with login, with a data table, and Supabase was hooked up automatically. That speed really caught me off guard.

But I care more about the second half. Lovable's loop is you type, it generates, you revise again—sounds smooth. The problem is the cost structure of this loop is asymmetric: the generation step costs almost nothing, while fine-tuning is what burns through your quota the most. I looked at the usage tips for the free tier, and the official docs themselves advise users to manually tweak padding, alignment, and that kind of stuff—don't let the AI do it, save your quota for code generation and debugging. The advice itself isn't wrong, but it happens to illustrate one thing: the most valuable part of the product is going from zero to something, the least valuable part is going from something to good—and the latter is the real workload.

In principle, code generated from natural language is an editable black box. The interface is given to you, but the design intent isn't, so to modify it you can only guess through repeated dialogue. That mood-tag feature the student made took seven or eight rounds of back-and-forth to get right.

For prototype validation, classroom demos, or building interfaces for non-technical collaborators, I recommend it. As the main tool for a production project, I don't recommend it.