Most thesis-writing guides follow a timeline: proposal in this month, first draft in that month. But when it's time to actually start, a timeline won't help you—what helps is first figuring out where you are right now: Do you have nothing at all? Have you finished the experiments and only have data? Or has your draft been sent back by your advisor?
Different positions call for completely different approaches. This article doesn't follow the calendar; it splits into three routes based on "what you have in hand." The tool is MuchSci's Paper Sailor, and the three routes correspond to the three ways it helps you start. You can first find the one that fits you, then read the route for your situation below.
First, remember two rules that apply throughout
No matter which route you take, MuchSci's Paper Sailor keeps to two iron rules:
1. Never fabricate—if there is no real data, do not write the experiment/results sections; it only gives you a list of "what's missing"; target-value placeholders are never filled with numbers on their own; references are used only as real entries returned by retrieval/database import.
1. The language version is not decided for you—before delivering a Chinese or English draft, it will ask you to choose once, and the output directory and file names will change accordingly.
For undergraduates, the first rule is especially important: the worst way a thesis can go wrong is for nonexistent data to appear in the results chapter. This rule effectively blocks the road most prone to accidents.
Figure 1 | Two iron rules that apply throughout: never fabricate / the language version is not decided for you.
Route A: Starting from zero—you have nothing yet, only a broad direction
Suitable for: those who have just been assigned a topic and whose advisor has only given a direction.
1. Set the direction: hand your vague idea to it to generate idea cards, refine the research question, and write a draft Proposal. The output of this step is not a "thesis," but something you can take to discuss with your advisor—align on the direction before you start writing.
1. Build a literature library: multi-source search (arXiv / Semantic Scholar / MCP services for deduplication) → automatically download PDFs and store metadata → generate a literature review → citation management (Bib/RIS/three-way verification/authenticity checks). Pay attention to the last part: specifically verify the authenticity of references and eliminate hallucinated references—the references in your proposal must stand up to your advisor's spot checks.
1. Produce a writing blueprint: generate a thesis outline, build the chapter skeleton, and mark which chapters do not depend on data (can be written first) and which must have real results (marked data_needed, waiting for experiments).
The key action in Route A is the third step: there are several months between proposal and defense, and the blueprint first separates "what can be written" from "what must wait," so you can write the introduction and review first instead of idly waiting for experiments.
Figure 2 | The three major steps of Route A: set direction → build literature library → produce writing blueprint.
Route B: You already have data—experiments are done, but not a word of the thesis is written
Suitable for: those whose data is sitting in Excel/CSV and who only realize in March that it's time to start writing.
Entry B starts from the data: give it your data files (CSV/Excel/JSON), and it automatically profiles them with pandas—fields, distributions, missingness; then performs data-gap mapping against the thesis outline and generates an action list: where data is sufficient, where it is missing, and what needs to be supplemented first.
After that comes iterative data supplementation: supplement whatever is missing, profile again after supplementing, and do not enter writing until the gaps are cleared. The greatest value of this route is that it eliminates the problem of "not being able to write the results chapter" before you start writing—missing data is not scary; discovering missing data only after writing the results chapter is what's scary.
Figure 3 | Route B loop: upload → profile → gap mapping → action list → supplement data.
Route C: You already have a draft—half written, sent back by your advisor, or almost finalized
Suitable for: actually, this is the largest group.
Entry C edits directly, à la carte:
· Polishing (removing AI flavor): a rule engine detects AI-like style—boilerplate transitions, empty strong assertions, translationese, etc.—and produces an AI-flavor index report; high items are rewritten one by one until they reach zero. Schools are increasingly checking for AIGC traces, so it is recommended to do this step before submission.
· Proofreading: a state machine of up to 3 rounds of proofreading + locking, with automatic backup for each round. When your advisor asks you to revert to the previous version, you'll be grateful for this backup.
· Add citations: citation management and authenticity verification, running through references of "unknown origin."
· Reply to reviewers: a response letter to review comments (item by item: category + action + manuscript location + response text). Undergraduates don't deal with peer review, but your advisor's revision comments can be responded to item by item in this format—what was changed and where, all clear in one table.
Figure 4 | Route C revision menu: remove AI flavor / proofread / add citations / respond item by item.
The three routes ultimately converge in the same place
No matter which entry point you come in through, once you enter the writing execution stage, the human-machine division of labor is the same:
· You are responsible for: real experiments, real data, and final decisions;
· It is responsible for: writing the main text, data analysis and subset filtering, and publication-quality figures (8 chart types + multi-panel assembly, SVG vector + journal column widths such as Nature 89mm).
For undergraduates, the figure feature is an unexpected bargain: journal column widths may not be useful, but SVG vector images stay sharp when enlarged, and are clear when pulled directly into defense slides.