When writing a review, the most hidden pitfall isn’t too few references—it’s “citing by feel”: when you reach a point, you think, “It seems like some paper discussed this,” and cite it from memory. If you get it right, it’s luck; if you get it wrong, it’s an accident. The lifeline of a review is that every sentence can be traced to a specific location in a specific reference.
This article breaks the reference work for a review into three steps, with the goal of building a “material pool”: every item in the pool has the original text, a source, and is ready to use whenever needed. The tool used here is MuchSci’s paper search assistant as an example—it works with your local paper library, and its answers are based only on real references in the library, with source chapters noted. This is exactly the form a material pool needs, and it can be seen as a graduate-student AI assistant that guards your reference library.
Figure 1 | Three steps for the material pool: define the scope, pull the list, deep-read.
Step 1: Define the scope—draw the boundaries of the pool
Before you start writing a review, first answer “which references in this field must be in the pool?” Tell it “help me find a few papers about XXX,” and semantic search finds the most relevant references by research topic rather than only matching title wording. Two rounds of search are better than one: use topic words in the first round, then switch to synonymous expressions in the second—semantic search is tolerant of rephrasing, and asking in different ways can fish out what the first batch missed.
Add filters at the same time: use the year range to lock down timeliness (work from the last three years cannot be missing from a field review), and use citation count to lock down weight (milestone older references should be collected separately in their own category).
Step 2: Pull the list—make the pool transparent
The pool must be transparent—don’t rely on memory to remember “roughly what’s in it.” Regularly ask: “What papers are in the database?” The paper list query in MuchSci’s paper search assistant lists all papers currently stored in the library.
Do three small things with this list: label each paper with a role (mainline reference / comparison reference / background reference); when you find an obvious missing corner, go back to Step 1 and search again with different keywords; and move what you’re sure you won’t use out of your attention—the value of a material pool lies in “every item in it is usable.”
Step 3: Deep-read—turn “citations” into “material you can actually use”
There are three types of material that truly carry a review when you write it, and the deep analysis feature supplies them category by category:
· Paragraph summary—for paraphrased sentences in a review such as “XX et al. approached from perspective A and proposed...”, paragraph summaries ensure the paraphrase does not distort the original;
· Proper nouns—the precise definitions of field terms; the source for the “term explanation” when a term first appears in the review is right here;
· Chart analysis—compare the experimental charts in various references; sentences with figures and evidence such as “the comparative experiment in Table 2 shows...” are the hard currency of a review.
Bottom line: two iron rules for the material pool
Its “Important Notes” are worth pasting verbatim at your desk: if the library has relevant content, quote the original text and note the source paper and chapter; if the library does not, explicitly say “no relevant information was found in the current paper library,” and never use training knowledge to fabricate an answer.
For a review, the second rule is even more valuable than the first—it means that in this material pool there is no “fabricated content that looks like a reference.” The kind of accident you fear most when writing a review—“it was cited, but no such reference can be found”—does not exist from the source under this rule.
Figure 2 | The two iron rules of the material pool.
On the day the review is finished, every item in the material pool should either have been used or explicitly discarded—there is no third state. Once you achieve this, “citing by feel” will completely retire from your writing process.