I am working on the same problem. My approach deals with it by sometimes commenting things out with // … or # … or whatever in the context window, but leaving comments that generated the code there.

One interesting thing, it seems to work better to have the model figure this out in a way it understands, then have it use that. Still trying related experiments.

I’m eager to read your work in more detail. I haven’t documented much lately—everything is moving so quickly, mostly just tryin to keep tryin, before my last idea is someone else’s deployed product.

I still highly value our hacker collaborations, I think they’re essential.