Seems everybody is rediscovering optimization? I see the LLM as a single step into a gradient direction. You can bend the surface a bit with your prompt, but the main influence are the weights. You can not really control the "step width" of that single shot answer. Optimization theory tells us to iterate and that it can become unstable under certain conditions. What you need to do for each step is to re-assess the loss function (how far are you from your goal). If you stop without, it is basically a gamble if that single shot is good or not.

While this is a pretty layman perspective, it helps me understand how to improve for now ok enough.