My advisor asked me to submit a first draft of a literature review. I “playfully” submitted two versions: one generated directly by WorkBuddy, and one produced with MuchSci through the full workflow from search to final draft. Below are the two first drafts and a record of my advisor’s comments.
Version A: Generated by WorkBuddy
The writing flows smoothly, the logic is sound, and reading it even moved me a little. My advisor’s comments came back quickly—the first was pinned to the third paragraph: “This 2019 paper—how come I’ve never read it? Go check.” I checked: no such paper. The second was pinned to a data point: “This 47%—where is it from?” There was no source. The third hit hardest: “Where did the references come from?”—they came from the model’s memory, and memory fabricates.
Version B: Generated by MuchSci
The approach was completely different: first conduct real searches in the 180 million paper Crossref database, break down the core points of each paper, and only draft after the direction is confirmed; behind every point is a real, existing reference, with sources annotated one by one and clickable for verification. This time my advisor’s comment was only one line: “This version is presentable; revise in this direction.”
Figure 16-1 Paper search assistant: 180 million Crossref papers, with core points broken down paper by paperTo be fair, Version A’s writing is not bad—WorkBuddy’s language organization has always been solid. The problem lies in the step before writing: it cannot guarantee that something is “real.” The first hurdle for a literature review is not the writing but whether the bibliography can withstand being clicked open.
Figure 16-2 The two origins of a literature review version: a four-dimension comparisonReverse the workflow—search first, then extract key points, then draft, and finally verify item by item—that is the reliable production line for a literature review. Writing tools are responsible for expression; source tools are responsible for truth. It is best not to expect the same model to do both from memory.