Seven practices I actually use when writing papers
Not a methodology—just seven moves I've kept using after two years of constant revision. For each, I'll say what to do and why.
1. Write the research question as one sentence before you start searching
What to do: Before searching, write down in one complete sentence what problem I want to solve, what the research object is, and where the boundaries lie. Then hand that sentence to a tool to break down the search dimensions, instead of repeatedly piecing keywords together in the search box myself.
Why: When keyword combinations don't produce results, you can't tell whether "no one has really worked on this direction" or "you didn't think of the word others would use." This is the trap beginners fall into most easily—mistaking a failed search for a research gap.
I use MuchSci's Paper Search Assistant: the literature comes from Crossref's 180 million papers, supports exact or fuzzy search by title, author, and DOI, and can also filter by year and citation count. When there's a match, it gives the DOI and automatically fills in author, journal, year, volume, issue, and page numbers—the citation information—so you don't have to copy it by hand.
2. Build a literature ledger with fixed fields
What to do: A table, one row per paper. Just eight fields: number, author-year, research question, method, main conclusion, limitations, whether cited, and which section of the paper it's cited in.
Why: If you don't build a table, you read thirty papers and end up citing five, and the only trace of the other twenty-five is "I think I've seen it." With a table, when you write the literature review, reorder it by theme and the structure emerges on its own; when writing the methods section, you can immediately find "how others explained the parameters for a similar method."
After a search match, have the tool break down each paper's research question, method, main conclusions, and limitations. These are exactly the four hardest columns to fill in the ledger. At the starting stage, filling in the table is far more reliable than "writing from memory after reading."
3. Read literature on two levels—don't use one posture for every article
What to do:
· Reviews and background—the goal is to figure out where the field has gotten to, so read quickly.
· Original papers strongly related to the topic—read the original text, focusing on methods, data, and supplementary materials.
Why: Using the same approach for both types of literature means you can't go fast where you should, and you skim past where you should be detailed. In the past I "read a lot," but in fact I didn't read either type properly.
4. Use translations only to build understanding; always verify facts against the original
What to do: On the first pass, use Paper Translation's paragraph-level side-by-side reading to read through and build understanding (the translation and original always stay aligned). On the second pass, check methods, parameters, values, and chemical formulas against the original. The role of translation is: bidirectional Chinese-English academic translation, consistent formulas and terminology, and paragraph-level side-by-side reading for entire PDFs.
Why: The cost of misreading one value in a translation is far higher than the cost of verification. In my first year of grad school, I misread an equivalent number and wasted three days.
One red line: A translation is not a factual basis.
5. Keep a separate glossary; do not use Chinese translated names as search terms
What to do: A "Chinese translated name + original English name" reference table, used consistently throughout.
Why: If the same concept appears in two translations in the main text, reviewers will spot it at a glance. Also, don't search databases with Chinese translated names—you won't find anything, and it's easy to be criticized for it; the original English name and DOI are what you need to carry with you at all times.
6. Archive as soon as you finish reading—don't let materials scatter in your Downloads folder
What to do: After finishing a close read of a paper, archive the translation and key points together. Put them in the Personal Knowledge Base (supports import of PDF, DOCX, PPTX, XLSX, TXT, MD, and other formats; personal use by default, not shared). For things from mixed sources—handwritten notes, whiteboard photos, files exported from other software—use the personal database layer, which supports more sources and can also take in reports generated by AI note-taking.
Why: Two or three years later, when you start a new project, you'll still need these. Scattered materials amount to no accumulation. If you often attend group meetings and talks, AI Note-Taking supports real-time recording translation, AI summaries, translation editing, and report generation. After recording, you get something you can listen to, view, and archive, and you can keep it together with your literature notes.
7. Before you start writing, talk it through once
What to do: Organize the structure into an outline, generate a first draft of slides (the Office Assistant Agent can produce editable PPT, Word, and Excel from a one-sentence description of your needs; for proposal scenarios there are templates—enter the project title, research question, and preliminary ideas, and it outputs a draft containing the research background, current status, methods, and technical route structure), then ask yourself while looking at those few pages: if I only present this, does the main line hold up?
Why: If you can't explain something clearly to a person, writing it down will be a mess too. When reading your own draft, you slide along with it; when speaking, you can't fool yourself.
Three boundaries must be made clear
· The viewpoints, arguments, and conclusions must be your own. Search, translation, organization, formatting, and generating a draft report can be handed to tools; what your research turns out to be cannot. Before submission or defense, confirm for yourself the target journal's and school's rules on AI use and plagiarism checks—this responsibility lies with the user.
· Fluent sentences do not mean the argument holds. When reviewer comments say "language needs improvement," in most cases the real problem isn't the language; it's that what you wanted to say wasn't clear in the first place.
· Calculate the cost yourself and make your own confidentiality judgment. These tools charge by usage (MuchSci uses credits; AI chat, paper analysis, and document generation deduct credits based on the model actually called and the processing volume, not a fixed charge per question; free registration gives 1,980 credits as a trial quota). Whether to upload unpublished experimental results, partners' sensitive data, or undrafted materials is your own decision. One thing I've confirmed: uploaded papers, data, and recordings are not used for model training, and encrypted transmission and access control are used.
·
Of the seven moves, the first four are searching and reading; the last three are consolidation and output. Once the earlier steps go smoothly, writing follows naturally.