That seems to be the core disagreement between the skeptics and the early adopters: what's the reasonable extent of usefulness any given piece of code is meant to have? Until quite recently, the extent to which your project "worked" was implicitly tied to your skill, understanding, comprehension, experience; to your ability to write code that is high-quality, easy to maintain, inspect, modify. LLM's completely decouple the one from the other. You no longer need to care about the quality, nor think of the maintainability, nor concern yourself with the ease-of-change and/or how brittle your codebase becomes after each and every new prompt. At least it "works", doesn't it?
Simply because it happens to "work" today, doesn't mean you'll require the exact same functionality of it come tomorrow. The requirements will change. What you want your piece of software to do will change. The way you want your piece of software to work will change. Unless you invest your time beforehand into understanding, maintenance, and quality of your codebase; unless you keep your finger on the pulse and avoid copy-pasting and/or trusting whatever output an LLM churns out for you, one merry day you won't be able to comprehend - much less control - the way it works, before you're no longer able to ensure it reliably works, before it stops working altogether and you're back to square zero trying to retrace and/or regenerate the whole codebase from scratch. Hopefully, you'll have no prod and no clients by then; otherwise you'll find yourself out of business entirely.
For this use case, I personally find LLM's perfectly viable. It's another piece of technology, which just like the rest of the pieces of technology that came before it, can only amplify what is already there. You've already laid the foundation for your project beforehand. You might have had some trouble recalling or understanding the choices you made back then, but you didn't start from scratch. You asked your LLM to analyze and supplement your own work with a few minor changes. You didn't generate the whole thing from scratch, outsourcing the entirety of the "how" and "what" to a server running a bunch of GPU's throwing around vectors and matrices with not a care in the world.
For the fun of it, try to recreate the entire project completely from nothing, by simply acting as the "project manager" for your token transformer. Then compare the choices it made to the way you approached it from the beginning. Who knows, the results might even surprise me.