You still need to prompt GPT-3 with something. Stacking two systems won't improve the outputs if your original input lacks precision. Garbage in, garbage out. Granted, GPT-3 requires way less precision to generate something useful, but you still need to pass the proper context.

I think you're right, though. It's hard to speculate how these things would turn out, but I think it's reasonable to assume that we will still need some relatively sophisticated human input to produce the best results out of these systems.