After the revision comments came back, I changed several habits in writing papers
A retrospective on paper writing: from the day I received Major Revision, I changed my approach to four stages.
I read that email three times before it sank in—Major Revision, four comments.
My first feeling wasn't discouragement; it was "why are they all about this?" None of the four comments questioned my experiments. They were all about other things: the literature review was incomplete, the methods were unclear, the citation format was a mess, and the language needed polishing.
In other words, the manuscript was sent back not because the research lost, but because the writing process did. During the month of revision, I changed how I handled these stages. I'll go through them in the order of the four comments.
Regarding the first comment: the literature review "omitted important recent work in the field"
This one stung the most because it was true.
My approach to literature at the time was: a few papers passed down from senior labmates + whatever I casually searched by keyword myself. The literature review read like a pile of loose sand—no tracking log, no concept of coverage. The two or three papers the reviewer pointed out—I truly hadn't read a single one.
During the revision, I did two things:
First, I changed the search method from "stringing keywords together" to "stating the research question clearly." Before, I repeatedly tried keyword combinations in the search box and gave up if they didn't work; now I describe what I want to ask in one sentence—for example, "Has the effect of X treatment on Y performance been verified?"—and let the tool break down the search dimensions. I used MuchSci's Paper Search Assistant. Its literature source is Crossref's 180 million papers, and it supports exact or fuzzy search by title, author, and DOI, as well as filtering by year and citation count. When a match is found, it provides the DOI and automatically fills in citation information such as authors, journal, year, volume, issue, and page numbers.
Second, I had each paper first "report its own conclusions," and then decided whether to read it. It breaks down each paper's research question, methods, main conclusions, and limitations. The greatest value of this step for me was coverage judgment: before, I could only judge "does this paper look like my topic"; now I can judge "to what extent has the question in this direction been answered." Literature reviews are often incomplete not because you haven't found things, but because you don't know which piece is still missing.
Regarding the second comment: the methods section "lacks reproducibility"
This one surprised me, because the methods section was the part I wrote most "smoothly"—smooth because I wrote it quickly, and the cost of writing quickly was vagueness.
The reviewer specifically pointed out two places: the value range of one parameter wasn't clearly stated, and the order of one processing step was ambiguous. I went back and checked, and the problem was that when reading the literature I only read the main text—many details were actually in the supplementary materials or footnotes in the methods subsection of the original paper. At the time, I was reading the gist in Chinese and didn't dig further.
During the revision, I changed the rhythm of reading literature: first read through the translation to build understanding, then check the methods, parameters, and data against the original text. I now consistently use a method that supports paragraph-level side-by-side reading—the translation and original text always line up, so when I reach a key paragraph I can verify it right there without searching back and forth between two files. I used MuchSci's Paper Translation. Its positioning is stated very directly: bidirectional Chinese-English academic translation, preservation of formula and terminology consistency, and paragraph-level side-by-side reading for entire PDFs.
Following this, I also stumbled on a byproduct: before writing the methods section, reread the methods paragraphs of similar papers in one batch. When writing your own, you may feel "there's nothing much to say in this section"; after reading five similar methods sections, you'll realize others specified seven or eight details you didn't write. This isn't a writing problem; it's insufficient input.
Regarding the third comment: the references "contain errors in format and content"
This one was entirely my fault, and it best illustrates "why these things shouldn't be carried by human attention."
Before the revision, I maintained references manually: copy the title, search for the author, type the journal name, count volume/issue/page numbers, and enter them into Word one by one. With over a hundred entries, errors were inevitable—one journal name had the wrong abbreviation, one year was off by a year, and there were two duplicate citations.
Now the process is: all citations take shape during the search stage. When a paper is matched, it gives the DOI and automatically fills in citation information such as authors, journal, year, volume, issue, and page numbers. When exported, it's structured content item by item, and no longer goes through the "I type it by hand" step. At the same time, I also set up a terminology list (Chinese translation + English original term) to avoid the same concept appearing with two translations in the main text—reviewers are very sensitive to this kind of inconsistency.
Regarding the fourth comment: the language needs polishing
I don't want to expand too much on this one; I'll just make two points.
First, using a Chinese draft to develop English expression: I first clarify the logic in Chinese, then move to English. In this process, I use bidirectional Chinese-English translation to go through it once, to see whether the sentence structure holds up. It's faster than forcing myself to stare at a blank page, but I still go through the submitted manuscript sentence by sentence myself.
Second, don't expect "language polishing" to solve argumentation problems. The review comment said "language needs improvement," but what really gets a manuscript sent back is often that what you want to say isn't made clear—language is only the outermost layer.
One thing I solved along the way: not being able to explain it clearly at group meeting
During the revision, I noticed a related problem: when I presented this work at group meeting, I explained it more clearly than in the paper. The reason was that by the later stage of writing, I had burrowed into the details and lost the ability to explain the main thread clearly.
Later I developed a habit: use a presentation framework to check the paper's structure. Specifically, I organize the paper's structure into an outline, have it generate a first draft of a PPT (MuchSci's Office Assistant Agent can produce editable PPT, Word, and Excel from a one-sentence description of your needs; it also provides templates for proposal scenarios—input the project title, research question, and preliminary ideas, and it outputs a draft containing the research background, current status, methods, and technical route structure), and then ask myself while looking at this draft: if I only present these few pages, does the main thread hold up?
Where it doesn't hold up is where the paper's structure has problems. This makes it easier to spot problems than repeatedly reading your own manuscript.
A few not-so-pleasant words
The revision made me realize that the line between "using tools" and "academic integrity" must be drawn clearly by yourself, so let me state it plainly here:
First, the opinions, arguments, and conclusions in the main text must be your own. Searching, translating, organizing, formatting, and generating a draft presentation—these can be done; what the research is written up as cannot be handed to anyone. This isn't a limitation of tools; it's an academic bottom line. Before submission, also confirm the target journal's statement on AI use and plagiarism-check requirements—this responsibility is yours, not the tool's.
Second, neither translations nor abstracts can serve as factual evidence. Parameters, numerical values, chemical formulas, and conclusions must be based on the original text. To this day, whenever I read any paper I'm going to cite, I go back to the original text and verify it.
Third, it has a cost structure. MuchSci measures usage by credits; AI chat, paper analysis, and document generation deduct credits based on the model actually invoked and the amount processed, not a fixed charge per question. Free registration gives 1,980 credits as a trial allowance, and monthly plans are 9.9 / 19.8 / 29.8. My suggestion is to try it on the paper you're currently writing—its value shows on your own manuscript, not in someone else's review.
Fourth, I've confirmed data security. Uploaded papers, data, and recordings are not used for model training, and encrypted transmission and access control are used. When a manuscript is under submission and data hasn't been published, this matters more than a few extra items on a feature list.
This revision ended up taking three weeks, much shorter than it took to write the first draft—because during those two and a half weeks, what I did was basically "make up for the laziness I had taken before."
The real efficiency of paper writing never lies in how fast you type, but in whether several stages happen in the right order: first finish reading what should be read, then verify clearly what should be verified, and only then write. If the order is wrong, what you write will sooner or later need rework—except the cost of rework is usually discovered for you by the reviewer.