Is the question worth asking, and is it new?
The first pass is about contribution. Does this paper address something the field needs answered, and does it add to what is already known? A technically flawless study answering a settled question is still not publishable.
Reviewers check novelty against their own knowledge of the literature and against the paper's own citations. A manuscript that does not engage with obviously relevant recent work invites the suspicion that the authors do not know the field.
This is where framing does real work. Reviewers are reading quickly at this stage, and a contribution that is not made explicit in the abstract and introduction may simply not register.
The first question is whether the paper adds anything. A sound study answering a settled question still does not pass.
Can the methods support the claims?
The core of most reviews. Does the design answer the question asked? Are the controls appropriate, the sample adequate for the analysis, the statistics matched to the data structure?
Reviewers look hardest at the gap between what was measured and what is concluded. Overreach — a correlational design supporting a causal claim, a result from one setting generalised to all — is probably the single most common substantive criticism.
They also check reproducibility. Is the method described in enough detail to be repeated? Are data and code available? A reviewer who cannot tell exactly what was done cannot endorse the result.
Is it presented so it can be assessed?
Structure, clarity, figures, and whether the results shown actually support the text. Reviewers are not copy-editors, but writing that obscures the argument becomes a substantive problem — an unclear paper reads as unclear thinking, and the report often comes back as "the contribution is not clear" rather than "the writing is not clear".
Figures matter more than authors expect. A reviewer frequently reads the figures before the results text, and a figure that cannot be interpreted without hunting through the caption is a genuine obstacle.
Finally the integrity checks: are ethics approvals stated where needed, are conflicts declared, are the references real and correctly attributed.
Writing the review itself
Separate what is fatal from what is fixable, and say which is which. A list of twenty comments with no indication of weight leaves the editor to guess, and guessing is what they asked you to prevent.
Be specific and locate every point — section, line, figure. "The statistics are inappropriate" is not actionable; "the paired test in section 3.2 assumes independence that the repeated-measures design violates" is.
Write as though the author will read it, because they will. Criticise the work, not the people, and note what the paper does well — a review that is only negative is less useful to an editor weighing a revision.
And give the editor a clear recommendation, with reasoning that matches the comments. A glowing report ending in "reject" tells them nothing.