I recently finished a full pass on a review article in my field, from outline to final draft. The biggest payoff wasn't the manuscript itself — it was a set of practices left behind after hitting every pitfall. Each point below comes with a real case.
Framework First, Writing Second
Before writing, I pulled five or six reviews widely regarded as excellent in the field and laid their chapter structures side by side. I wasn't comparing length; I was looking for shared patterns: whenever a review covered both synthesis and properties, synthesis always came first; whenever there was mechanism content, it always sat before the case studies, serving as an analytical tool.
Once I had the common patterns, I compared my own framework against them. Being different is an asset, not a violation — my manuscript has a standalone chapter on evaluation methods, which is rare in this field and worth keeping. But the placement of a distinctive chapter has to hold up: if it breaks the main thread of how it's made, what it does, and how it's tuned, it has to move.
When I review a framework, I check three things: whether the chapter order breaks the main thread, whether the classification logic stays on one level, and whether there's structural redundancy — the same case discussed once in the methods chapter and again in the mechanism chapter is redundancy.
Citation Rules: Recent, Rigorous, Primary, Traceable
After this manuscript, my citation rules grew from three to four. The new one is "rigorous":
- Recent: cite literature from 2021 onward whenever possible. A review's value is in covering progress; older papers are for established facts only.
- Rigorous: key claims must rest on solid papers in the field. My own rule of thumb: "You are only as good as the papers you cite." For citations supporting core conclusions, neither journal bar nor experimental completeness is negotiable. If data is missing, log it as a to-do — never fill the gap with a weak source. That's non-negotiable.
- Primary: cite experimental papers, not reviews citing reviews. A review's chain of evidence must land on original research.
- Traceable: data must be traceable. Data from an open-access review can be used, but only after tracing one step back to its original source. If no reliable source can be found, downgrade to a qualitative statement or drop it.
How did "rigorous" get added? I learned the hard way. In one section on doping-based tuning, I was missing a comparison dataset and casually cited an open-access paper to fill the gap. A later check showed its experimental design couldn't support my sentence at all. I deleted it, logged a to-do for the now-unsupported sentence, and later patched it with combined evidence from two solid papers. Better to leave a gap than fill it with a weak source.
A companion practice: scout literature at the 100-paper scale. For each section's literature pool, I set my target at 100 papers — cast a wide net, then filter. Stopping at a dozen means assuming the first dozen are the best. They probably aren't. From 100, I read 20 closely and keep the rest as background exclusions — and the exclusion process itself is writing material: if you can say in one sentence why each paper wasn't used, you genuinely know the field.
Citation Accidents and the Full Audit
This was the most painful lesson of the manuscript. Midway through writing, I hit three citation accidents in a row:
First, a citation to Alafeef with completely wrong volume, issue, and page numbers — worse, the paper itself is a review, but I had cited it as a source of experimental data. The correct reference is Small 2024, 20, e2303937, and it can only be cited as a review, never as an experimental source.
Second, two wrong article numbers — authors and years right, article numbers wrong, pointing to papers that don't exist.
Third, a 2023 paper by Sari et al. (*Carbon Resources Conversion*) that I had tagged "compliant," until I realized the journal is a fully open-access KeAi title. A different failure from the doping-section story above: there the experimental design couldn't carry the claim; here I misjudged the journal itself — by my own citation rules, it never belonged in a key-claim position.
The common root cause, in one sentence: "reusing previously verified references" is not the same as actually verifying them. Many entries were jotted down early, felt verified at the time, and were were reused from then on without anyone checking them word by word a second time.
So my process now has a hard step: a full reference audit before the manuscript is done, checking three things — placement (does this citation actually support this sentence?), relevance (is the cited work really about this context?), and authenticity (search the exact title in quotes, to confirm the paper exists and the details are right). No shortcuts; one entry at a time.
Citation Number Discipline
A small, specific, extremely common bad habit: a citation number [N] is only an insertion marker. Never use it as a subject, object, or noun phrase.
Wrong: "The study of XX et al.[3] shows that doping improves activity." The [3] gets wedged into the subject, reading like "XX et al. 3" is a person.
Right: "The study of XX et al. shows that doping improves activity [3]." The number always trails the sentence it supports. Trivial — but across hundreds of instances, inconsistency looks unprofessional.
Prose Standards: Nominal, Parallel, Rhetoric-Free
All headings are nominal and parallel — that's what keeps a review looking professional. The introduction gets no subheadings; one continuous stretch.
For the main text I run an eight-point self-check: no metaphors, slogans, or scare quotes; no unevidenced evaluative adjectives; no archaic or stilted flourishes or draft-stage phrasing; impersonal academic statements as the default; standard academic sentence patterns for definitions; parallel, factual phrasing for limitations; mechanisms stated as logical steps; endings flat and direct.
Two additions from this manuscript. First, language review covers every paragraph — no sampling. Sampling is gambling; the problem always hides in the paragraph you skipped. I go through each one, and on the re-read I focus on whether the edits introduced new problems.
Second, no vague time phrases like "in recent years" in the main text. Either give an explicit year range or state the fact directly. "In recent years, doping strategies have attracted wide attention" — which years, counting from when? Change it to "Since 2021" or just state the fact. Vague words get cut.
Two recurring awkward phrasings I caught and turned into permanent checklist items. One is misplaced subjects: "The evidence for X comes from the work of Y et al." — evidence doesn't "come from" a paper. Rewrite as "The work of Y et al. provides evidence for X." Put the subject back where it belongs. The other is homemade jargon: I once coined a term for "samples treated for different durations" that only I understood. Replaced with plain words.
One more lesson paid for in blood: no section pointers in the main text. "See §3.2" is a construction mark from the detailed-outline stage; it must never leak into the prose. Paragraphs connect through content logic — definitional openings, transitions that explicitly refer back — never through numbering.
Table Discipline: Column Headers Are Content Boundaries
Tables in a review rot the fastest, because a loose hand stuffs everything in while filling cells. My discipline: column headers are content boundaries.
If the column is "doping system," it holds only doping elements (N, N/S, N/P) — no preparation methods, precursors, or performance numbers squeezed in. If there's no column for a piece of information, it doesn't go in.
Two companion rules: performance data with an incomplete evidence chain doesn't enter the table — putting it there endorses it; single-source data stays qualitative, no numbers — a number with one source, written into a review, is where error propagation starts.
Removing the Scaffolding
"Anchor," "flagship section," "benchmarking" — I used these words constantly while writing. They're scaffolding. At delivery, the scaffolding comes down. Readers should see the article's own logic, not my writing process.
The test: hand the article to someone who wasn't involved in writing it. Can they follow the logic without me explaining "here I was following review X"?
Multi-Round Iteration: Review–Revise–Review, Faster Each Round
Depth in prose comes from iteration, not from a single pass. After finishing each part, I run fixed three-step rounds: independent review, revision, my final sign-off.
- Independent review: the latest draft is evaluated back-to-back against fixed standards — sharp, objective, evidence-based, no pleasantries, every criticism with its argument. Drafting and reviewing are separated; nobody reviews their own draft.
- Revision: revise point by point against the review report.
- My final sign-off: I decide whether the round is done or needs another pass.
The third step is where it compounds: besides signing off the draft, I usually add new yardsticks to the reviewers — each newly spotted problem becomes a new standard on the checklist. Standards accumulate, and each round starts stricter and sharper than the last, so I stop flagging the same class of problem twice. The awkward-phrasing and homemade-jargon checks mentioned above grew exactly this way.
A collaboration rule: state your position first, revise after the decision. When a collaborator or reviewer raises a point, I state my stance objectively first — agree or disagree, on what grounds, no pandering — and only revise after the call is made. Never revise first and explain later; then the other side never learns why you changed it, and the same question comes back next time.
Audit records go at the end of the finalized file, marked as internal material — never as separate files. A separate file creates a second source of versions that will eventually disagree; appended at the end, I can still look up three months later what changed in each version, why, and who signed off.
Dividing Labor Among Subagents
The review–revise–review loop above only runs because of a fixed division of labor among subagents, in three branches:
- Writing branch (serial): a drafting agent writes or compresses → an independent back-to-back reviewer evaluates → I revise → the user signs off, looping until done. Iron rule: the same draft is never both written and reviewed by the same agent.
- Materials branch (parallel): literature scouting, figure sourcing, and data verification each run on their own, nobody waiting on anyone else; I assemble everything once everything is in.
- Distillation branch (standalone): experience notes and method docs are drafted on a separate track and never pollute the main manuscript.
The user signs off every branch as the final gate.
Mentoring: Turning Experience into Deliverable Rhythm
Midway through this manuscript, I started handing the follow-up work to a student. The rhythm we found: each week's delivery = prose increment + a 3–5 page slide module + figures.
The key is how the slides are used. I require modular, jigsaw-style reporting: each week, 3–5 pages that go deep on exactly one module — one week a single carrier system, the next the recycling strategy. Rebuilding one giant deck every week is forbidden — that's pure layout labor producing zero new content. Modules are polished independently and snapped together at the end into the full talk.
I tell the student upfront where each deck goes, four destinations: polished mechanism schematics and classification diagrams become the manuscript's core figures; the full deck plus narration becomes the journal's video abstract; the finished deck is frozen as onboarding material for new group members; after acceptance, core pages are pulled out as outreach material for academic platforms. Build once, use four times.
My rule on AI use: AI may assist, but three things stay human — AI can search literature, sketch drafts, and format; but whether a citation is on-topic, whether key data was checked against the original paper, and whether you can explain a figure's logic yourself — those three are human. AI drafts, humans judge; whoever signs off owns the quality.
Conclusion
The order can't be reversed: framework first — writing before the framework is settled means rework later; then citation rules — sources for key claims fixed before drafting, gaps logged as to-dos; then prose, paragraph by paragraph; and every pitfall converted into a checklist item, standards accumulating, each round faster. Mentoring works the same way: rhythm, deliverables, and where each slide ends up — stated clearly before work begins.
← Back to Blog