Are we over-optimizing prompts instead of designing better thinking systems?
Most of us spend hours tweaking prompts — adding structure, constraints, examples — just to get slightly better outputs from AI tools.
But I’m starting to wonder: are we solving the wrong problem?
Instead of endlessly refining prompts, maybe the real leverage is in designing systems around the model — things like memory layers, feedback loops, or multi-step reasoning workflows.
For example:
-storing previous outputs and reusing them as context
-breaking tasks into smaller chained steps
-letting the model critique or refine its own answers
This shifts the focus from “how do I ask better?” to “how do I build a system that thinks better over time?”
Curious how others approach this — do you invest more in prompt engineering or in building structured workflows around AI?
I think there is another layer before prompts or systems.
Neither improves very much unless the system can observe the outcome of its own decisions.
A refined prompt without feedback is still a static prompt. A sophisticated workflow without meaningful signals is still repeating the same process.
That is why I find feedback loops more interesting than prompt engineering itself. They give the system something to learn from over time instead of simply executing better instructions.
For me, the real shift is not from prompts to systems. It is from systems that execute to systems that can observe.