Thompson et al. present a work where they:

-collate different pre-existing databases of fish stomach contents into a single one, including almost one million "predator' entries from the North Atlantic shelf and arctic seas  over the 1936-2020 time period;

-uniformize trait variables in the database (prey individual size, prey biomass, and prey density) by modeling missing entries using a predator-prey mass scaling predictive model

-develop a new categorization of predators (species X mass) into fish trophic guilds based on a broader diversity of trophic traits computed from the uniformized database

-apply the method to a survey dataset from the N-E Atlantic shelf.

The datasets are impressive, the methods are rich, and most are very well-detailed. The application example sufficiently illustrates the potential of the databases while not spoiling future works and investigations. Overall, the datasets could be of use to many academics as well as marine managers, and the overall approach could be inspiring for other biodiversity research based on trophic traits.

However, the manuscript needs to be repurposed to align the claimed objectives with the provided datasets and codes.

Suppose the purpose is to produce a uniform dataset of fish stomach contents, which, despite different protocols, enables the display of more information than solely prey occurrence, i.e., species and size-specific feeding traits. In that case, this purpose has been reached as both the codes (glmm) and the databases (before and after uniformization) are provided along with the manuscript, and the methods to simulate the missing data are provided with sufficient details. However, the end products do not fulfill the expectations of the title and claimed objectives of the paper's introduction (i.e., fish guilds). 

As claimed by the title and introduction, the purpose is to provide a new categorization of fish guilds based on a broader set of trophic traits (rather than just occurrence and biomass). In that case, the objectives and the provided datasets/codes are inconsistent.

The database for fish trophic guilds, which should be central, is provided for now only as a table in the supplementary material (while a csv is needed for this info to be reused). The codes to generate the clustering that could be used to update or check the robustness of guilds over other regions are not provided. The categorization efficiency based on feeding traits has been tested against previously developed approaches with similar objectives. Still, we are provided only with very few results, although this should be a major motive for the study. Why this new one is more robust is not sufficiently demonstrated (only through the distance to the centroid, for which we have very few results provided).

Making the manuscript a better fit for the latter point is an easy fix. The introduction should be more informative, not only to address beyond the very specialized marine fisheries scientists but also to better account for previous works on this topic. This is not the first attempt to provide an estimator for fish guilds. Still, it seems this one is expected to leverage databases to be more informative (by simulating the missing traits) rather than by working with the minimally common data in the heterogeneous databases (occurrences and biomass). It is only in the late discussion that previous attempts are acknowledged. 

Specific comments

Unfunctional web links l 95 and l138

paragraph 2.2 and throughout the ms: I found referring to prey and predator individual mass as "body size" confusing. This is probably a common term, but it is improper.

paragraph 2.2 the glmm explanation needs to be clarified: what are the predictors and the response formula? random and fixed terms? all is here but it is only when I checked the formula on the GitHub that I got a clear idea. the glmm formula should be written in plain, and the text should be polished to be better understandable.

Fig 2: many misspelling in it; also, what is the pink category?

thoughout the mansucript, a zooplanktivorus fish is a predator, even if found in another fish's stomach. this is an example of the marine jargon's problem of the paper. the "Prey "and "predator" terms should be defined upfront: it would be much more transparent for the reader if, from the beginning, preys were defined as items found in stomach contents and predators as the sampled fish. 

How are uncertainties accounted for in the predictions?

Fig 3: higher intercepts and slopes for the biggest fish-fish interactions (l313) cannot be observed in Fig 3.

Thompson et al. compile an impressive data set of stomach samples from several countries from the Northeast Atlantic shelf seas. They use this to provide empirical estimates of species- and size-specific feeding traits of fish. This is an interesting study that appears to be well thought out, and do not have any major technical concerns (number 3 might be the only one that requires some work).

Main comments and suggestions

1. Line 81: suggest defining 'feeding guild species richness', as it is not intuitive whether this refers to the species richness of the predators or of the prey. Line 83 suggests it is referring to prey species, but given a feeding guild is a group of predator species that feed on similar prey, the species richness of the guild would seem to be the number of predator species.

2. Lines 42 and 262 mention the work being reproducible. It is commendable that the the raw data appear to be fully available on the Cefas Data Hub website (in a 47 Mb .csv file!). I checked the GitHub site for the code (L193), and it appears to be well documented and usable by interested parties, which is commendable. Given this, I think it should be mentioned why the data (and therefore the analyses) stop in 2020. If the full workflow was reproducible then it should have been fairly easy to update the analyses. I appreciate that the collating of data into the big .csv file is likely hard to somewhat automate, given the different databases involved, and this might be the barrier to easily updating the analyses (and you have to cut the data off at some point in time). Or maybe all the data were simply not yet available. I'm thinking ahead about whether the

analyses could be updated to operationalise the work, given the suggestion of a candidate food web indicator (L501), which presumably would need to be continually updated if it were to be used in fisheries or ecosystem management.

3. I think the modelling to fill in missing data (L163-167) is sound, but was a sensitivity analysis conducted to compare results when using only data for which all variables are available? I did not see one but it would make the work more convincing. It would at least highlight the impact of the assumptions, and such results may depend on the relative differences between partial and complete data sets (and how those difference differ between areas or guilds).

4. Figure 2 is not referenced in the text, and should appear earlier on than it currently does. The red circles are not defined, nor are the arrows. I think the arrows represent flows of information, but the thin red arrows are a little unclear. Maybe the red circles should just be another step along the black arrows? The rest of the diagram is clearer.

5. L310: The sentence "There were significant differences in the log10 transformed scaling relationship between predator and prey body mass captured by the interactions between predator body mass and predator species and prey functional group combinations (Fig. 3)" is very unwieldly. I think it could be reduced to something like "There were significant differences in the log10 transformed scaling relationship *of* predator and prey body mass between the different predator-prey functional group combinations (Fig. 3)" The "captured by the interactions between predator body mass and predator species" is somewhat implicit and is not needed, unless I am missing the intent of the sentence. The Figure 3 caption is clearer.

6. Figure 3: panels are really small given the amount of detail; similarly for Figures 5 and 6. Also, it is a little unintuitive that prey is on the y-axis and predator on the x-axis, given what is being shown is the predator-prey mass ratio (and we normally think of a y-value being divided by an x-value). I guess I'm also used to thinking of predator-prey models where prey would be on the x-axis. I checked some predator-prey mass ratio papers and this does seem to be the usual way to plot these, but maybe just add a comment somewhere. L347 confirms that a higher PPRM is a lower point in Figure 3. It's a little confusing.

7. L320: How is the subset of 498 combinations chosen to be shown in the figure? This should be stated, as it is a very small subset of the full results.

8. L320: Suggest referring explicitly to the panels (and call the bottom one (m)), and explain earlier in the caption that (a)-(l) are ordered by prey size at maximum predator size (it's confusing to mention that when talking about panel (m)). And the grey points are only in (m), yes? Could be more explicit.

9. Figure 4. Again, labelling of panels might help, especially as the first three are a bit different to the second three (but on first reading, only four types of plot are described).

Minor comments:

L103: it might be clearer to change "Predators feeding on relatively small prey with high predator-prey mass ratios (PPMR) can dampen strong oscillatory dynamics" to "Predators feeding on relatively small prey yield high predator-prey mass ratios (PPMR) that can dampen strong oscillatory dynamics". The prey do not have high PPMRs.

L126 mentions extending to other areas such as the Northeast US continental shelf, but such data are already included.

L134: DAPSTOM needs defining, and I think a few other acronyms might also.

L135 and Table 1: Figure 1 shows data off the east coast of Labrador, Canada, but no Canadian data sources are mentioned, just Northeast US

continental shelf. This is not accurate if the term is meant to include the Canadian data.

L169: the model descriptions would help by also having a simple equation of the models (in word form is fine as is often done for such models), as it would summarise the approach.

L177 'were' should be 'where'

L381 and L382: 'lower in the food web' is a bit superfluous given the 'planktivore' reference, though maybe '(i.e., lower in the food web)' would

work. Similarly for 'higher in the food web'.

L381: The Celtic Seas and Bay of Biscay need defining or highlighting on a map.

L407 and likely elsewhere: Suggest hyphenating 'body size' when used as an adjective.

L408: Really it's just 'parts' of the Atlantic and Arctic Oceans, not the full extent (which is how it reads). L415 mention of 'Northeast Atlantic shelf seas' is more correct.

L491: "We make use" and "We then developed". Be consistent with tense.