From Proposal to Final Draft: Five Bottlenecks in Graduate Thesis Writing
An observational note on the graduate thesis writing process, recording the pitfalls I stepped into at several stages and how I later handed off some of the repetitive work.
When it comes to writing a graduation thesis, what truly breaks people has never been the act of "sitting down and typing."
Before writing, you know what you want to write; while writing, you know what you're saying—most of the time is actually spent elsewhere: two hours digging through databases, and the literature you find only barely touches your topic; you download thirty PDFs, but fewer than ten have actually had their abstracts read; you finally settle on a topic, and on proposal day your advisor asks, "Has anyone done this direction in the past few years?" and you freeze on the spot.
Later I broke these bottlenecks down and counted them. During the writing process, about five stages repeatedly drain people, and three of them are "purely mechanical labor" that has nothing to do with your intelligence.
1. Proposal Stage: Not Knowing Who Has Already Plowed This Field
The scariest thing about a research proposal is not an immature idea, but an outdated one.
At my first proposal, I offered an angle I thought was quite new. My advisor replied calmly: "There have been quite a few reviews on this topic in the past three years. Go back and sort through what already exists." The problem was—I didn't know what "already exists" meant. I had not searched systematically; I had only read a few scattered papers passed down by senior students.
In academic training, "understanding the current state of the field" is assumed to be something you know how to do, but no one has really taught you how to quickly build a map out of tens of thousands of papers. If you don't get past this stage, the methods and innovations that follow are all castles in the air.
2. Literature Search: If Your Search Terms Aren't Right, the Results Are All Noise
This was the stage where I got stuck the longest, and also the area where I changed the most afterward.
The problem with traditional search isn't that the databases are bad; it's that literature search itself is highly experience-dependent:
l What you want to ask is a research question ("Has the effect of X on Y ever been verified?"), but the database can only take keyword combinations;
l One word off in the search terms, and the results are two different worlds;
l Hundreds of results come out, and you have to judge relevance one title at a time;
l Once you pick one, you still have to manually fill in the author, journal, year, volume, issue, page numbers, DOI—one paper at a time.
An entire afternoon can disappear in the loop of "revise the search query → scroll through results → exclude → fill in citations."
Later I switched to describing the research question in natural language instead of forcing out a search query. The tool I used is called MuchSci. What its paper search assistant does is this: you state your research question clearly in one sentence, and it automatically breaks it down into search dimensions; the underlying literature source is Crossref's 180 million academic papers; it supports exact or fuzzy search by title, author, and DOI, and can also filter by year and citation count.
The two things most useful to me weren't fancy:
First, the "citation completion" step is eliminated. Enter a title or author, and it returns matching results and the corresponding DOI, then automatically completes citation information such as author, journal, year, volume, issue, and page numbers from Crossref and other data sources. What used to require typing into Word by hand and checking item by item becomes a single export. Anyone who has written a reference list knows that what this saves isn't time—it's the irritation of proofreading.
Second, it doesn't just give you "a list of items." It breaks down each paper's research question, methods, main conclusions, and limitations, and supports exporting key data. This means that during the screening stage, I don't have to actually finish reading every PDF before deciding whether to keep it—first look at the breakdown, then decide which ones are worth reading closely. Thirty papers become ten, ten become three, and the density of reading goes up.
To be clear: you still have to judge the search results yourself. The tool gives you a "cleaner candidate set," not a conclusion about "which papers you should cite." The quality of a literature review ultimately depends on whether you read carefully enough.
3. Reading Stage: Language and Time, Two Big Mountains
English makes up the bulk of the literature; that's a public fact. The pain of reading English literature has two layers: one is genuinely not understanding the terminology, and the other is understanding it but not being able to speed up—in the same two hours, you can get through ten pages in Chinese but only three in English.
My approach is layered: for anything strongly related to my topic, I read the original closely; for material only meant to understand background or confirm a certain conclusion, I let paper translation go through it first. MuchSci's translation supports 13 common languages (including Simplified and Traditional Chinese) and can handle docx, pdf, txt, and md formats.
By the way, the design logic of its "translation" feature is quite restrained: one-click submission, professional results, without too much extra packaging. My usage is simply "get the gist first, then go back to the original for key paragraphs." This combination is much faster than grinding through it, and cleaner than my old habit of cutting paragraphs and pasting them into a general translation tool.
4. Writing and Reporting: The Least Valuable but Unavoidable Work
In the middle of thesis writing, there is a plateau period where "the text moves very slowly." What actually interrupts your rhythm at this point is often not the content, but the supporting outputs around it:
You have to present at a group meeting, so you need to make a proposal or progress report PPT;
Your advisor wants an outline, so you need to extract a structured Word version from the thesis;
Data needs to be organized into tables, and Excel gets reworked over and over.
The characteristic of this work is: it's not hard, but each item takes up a whole block of your complete attention.
MuchSci's office assistant agent solves exactly this part—you describe your need in one sentence, and you get editable PPT, Word, and Excel deliverables. It has templates for proposal scenarios: enter the project title, research question, and preliminary ideas, and it outputs a draft PPT with structures such as research background, current status, methods, and technical route. I don't treat what it generates as a final draft, but it's very useful as a "starting point"—writing a structured page from scratch and revising a page that is already structured are completely different in psychological cost.
Another scenario worth mentioning: if you often attend lectures and group meetings, its AI note-taking supports real-time recording, summarization, and report generation. Record a piece of audio and you get something you can listen to, review, and archive. When writing literature notes and meeting minutes, it can save some of the cost of recalling.
5. Throughout the Whole Process: Making Existing Materials Truly "Usable"
This is something I care about more the longer I use it.
Graduate thesis writing is not creation from zero; it is taking one step forward on the foundation you have accumulated over the past two or three years. But in reality, these accumulations are scattered across: local folders, PDFs sent via WeChat, snippets in note-taking apps, and recording summaries left from the last meeting notes... Change to a new topic, and it's like going through everything all over again.
In MuchSci, I value the personal knowledge base and personal database layer more: the knowledge base supports importing PDF, DOCX, PPTX, XLSX, TXT, MD, and other formats, and by default it is personal and not shared (in team collaboration scenarios, a shared knowledge base can be configured in the backend); the personal database supports more miscellaneous sources and can collect files from multiple sources, including reports generated by AI note-taking. Its positioning is clear—let existing knowledge support new research.
Add to that "personalized answers"—it responds to your focus with your background and context, instead of starting from "what is a literature review" every time. This becomes most obvious after you've used it for two or three months.