Dear Authors,
Based on two thoughtful reviews and my own impressions, I suggest major revisions before resubmission for a subsequent review cycle. I agree with the reviewers that the results are quite equivocal about the relative merits of the tested approaches. This outcome is of course publishable, and possibly useful to the field, provided the authors draw conclusions that are consistent with the results. Thus, the possible conclusion that the methods adopted do not, in fact, robustly improve the baseline/reference forecasts, or even degrade them, must be given some consideration by the authors -- this seems true beyond the first month, at least. I also urge the authors to consider the suggestion regarding conditioning with temperature or other variables, as this is more skillfully predicted by the GCM, and may give more positive results. I realize that this would require significant work, however, and let the authors judge whether it would be possible for this paper.

In addition, I think that there are several curious features of the results that warrant further explanation. In particular, in Figure 6, I'd like the authors to give a more thorough analysis of the flipflop in skill within the first 15 days. A second suggestion is that the use of the last obs as a conditioning factor for the hist-based ensembles may not be a bad idea, but obscures the impact of the conditioning on other factors, which is of interest (and more comparable to the Sys4 based ensembles).

As both reviewers observed, neither the use of conditioning and of Sys4-driven leads to compelling skill upgrades from the standpoint seasonal prediction, yet this aspect is not one that could be improved without the work becoming an entirely different study, perhaps using different methods and data. The authors argue that the goals of the paper are not necessarily to provide a new and improved strategy for seasonal flow forecasting, but to highlight the impacts of the conditioning strategies on forecast qualities. I accept the paper because (a) I agree that there is value in such an analysis; (b) the demonstration of limited skill (though limited or negligible in some cases) achievable through these strategies may be useful information for the field; and (c) though not highlighted, the comparison between using model inputs directly from Sys4 versus Sys4 as a source of a conditioning factor is also interesting. The paper may have been strengthened by focusing not only on sharpness and reliability but also on attributes such as bias and resolution, which can strongly affect overall skill scores like the CRPSS. Nonetheless, the findings are clearly presented and should offer a useful data point for researchers and practitioners.

I am returning the paper only for minor clarifications and further proofreading.