In my experiments with recursive development, you would have to train the thread to understand prompt engineering as a whole not just it’s understanding of the task at hand.
Key example, you want it to spit back a set of pose prompts for images, but it doesn’t know what it doesn’t know, so all it will do initially is pull a list of all the things it knows are poses and repeat them.
If you take the time, and build up its understanding as a whole of what your process is, and why you need it, and give the context window enough time to form some coherence in it, you get significantly improved results as long as you don’t trip automation scripts in the model that revert it’s behavior.