I’m not sure what your end goal is and there are certainly use cases where you want free-form input and don’t care so much about the predictability of the data, annotations and experiments. But by design, Prodigy tries to limit the modifications made by the annotator to a few selected properties (answer, spans, label). There should always be an input (the data to be annotated) and an output (the input plus annotations). In the scenario you describe, this is different: you’d have nothing coming in, and arbitrary unpredictable data coming out.
This makes it super difficult to run predictable and structured experiments, so we’d usually advise against workflows like this. Instead, it’s often more efficient to move the unpredictable, free-form data collection out into a separate step: for example, start by collecting the data from the user and then annotate it with Prodigy.