I have actually very convincingly recreated a moderately complex 70s-era mainframe app by having an LLM reimplement it based on existing documentation and by accessing the textual user interface.

The biggest trick is that you need to spend 75% of your time designing and building very good verification tools (which you can do with help from the LLM), and having the LLM carefully trace as many paths as possible through the original application. This will be considerably harder for desktop apps unless you have access to something like an accessibility API that can faithfully capture and operate a GUI.

But in general, LLM performance is limited by how good your validation suite is, and whether you have scalable ways to convince yourself the software is correct.