The Writing Process

Knowing what a chapter should contain does not produce the chapter. This document is about mechanics: what order to write in, how to generate pages reliably, and how to use your advisor so that feedback arrives while it is still actionable.

Write in this order

Not chapter one to chapter eight. Write in order of how well you understand the content:

1. Design — during Phase 2, while you are designing
2. Related Work — weeks 8–12, while papers are fresh
3. Evaluation — as results arrive
4. Implementation — right after the system stabilizes
5. Background — late, because only now do you know what background is actually needed
6. Discussion and Limitations
7. Introduction — near the end, when you know what you are introducing
8. Conclusion
9. Abstract — last

Two reasons this ordering matters. The introduction and background can only be written well once you know what the thesis actually says, and writing them early guarantees a rewrite. And the design and related-work chapters are perishable in a way the others are not: rationale and rejected alternatives decay from memory within weeks, so capturing them contemporaneously saves real work.

Outline to the paragraph level before writing prose

The highest-leverage technique in this document.

Before writing a chapter, write a topic-sentence outline: the first sentence of every paragraph you intend to write, in order, and nothing else. A twenty-page chapter becomes perhaps sixty sentences on two pages.

This works for two reasons. Reading only the topic sentences in sequence exposes broken logic immediately — you can see that section 4.3 assumes something you establish in 4.5, or that two paragraphs make the same point. Fixing that at outline stage costs a minute; fixing it after the prose is written costs an afternoon and you will be reluctant to do it, because by then you are attached to the paragraphs.

It is also the right thing to send your advisor. Two pages of topic sentences gets you structural feedback in days. Twenty pages of prose gets you comma corrections in three weeks, because reviewers respond to what is in front of them, and finished-looking prose invites line editing rather than structural critique. Sending the outline first is how you get the expensive feedback early.

Once the outline is approved, writing the prose is mostly mechanical expansion, at maybe two to four paragraphs per hour.

Daily cadence

Five hundred words a day beats five thousand on Sunday, for the ordinary reason that the Sunday plan collapses on the third Sunday and the daily habit does not. At 500 words a day, five days a week, an 85-page thesis takes about seven weeks of drafting — which is exactly the Phase 4 budget.

Some practical points. Write at a fixed time, ideally when you are freshest; do not schedule writing for whatever hours remain after debugging, because there will not be any. Track words per day somewhere visible. And never edit while drafting — this is the rule people break most and it is the main cause of the two-hour session that produces one perfect paragraph. When you cannot find a word or a number, write \todo{} and move on. Drafting and editing use different mental modes and switching between them constantly is what makes writing feel impossible.

Leave a sentence unfinished at the end of each session. Starting is the expensive part, and resuming mid-sentence skips it entirely.

Revision in separate passes

Revise in passes, each pass looking for one class of problem. Trying to fix everything at once means fixing nothing well.

Content pass. Is every claim supported? Every figure discussed? Every \todo{} resolved? Anything promised in the introduction actually delivered?

Structure pass. Read only the section headings and topic sentences. Does the argument flow? Anything out of order or duplicated?

Paragraph pass. Does each paragraph make one point, stated in its first sentence?

Sentence pass. Now you may care about prose. Cut hedging and filler, prefer active voice where it is natural, break up sentences over about thirty words.

Consistency pass. Terminology used identically throughout — pick one of "policy", "predictor", "model" and stick to it. Notation consistent. Numbers in the abstract matching the numbers in chapter six, which is a genuinely common error after a late experiment rerun. Consistent tense: past for what you did, present for what the system does.

Reference pass. Every citation resolves, no [?] markers, every figure and table referenced from the text, no broken cross-references.

Proofread pass. Print it. You will catch things on paper that you have looked past on screen for weeks. Reading aloud catches a different class of error again.

Budget about a week per pass on a full draft, and note that revision is where the quality comes from — the Phase 5 budget is not padding.

Working with your advisor

Advisor management is a skill and it is worth being deliberate about, because latency in this loop is the thing most likely to compress your endgame.

Send outlines before prose, for the reasons above.

Send a weekly written update whether or not you meet, using templates/weekly-advisor-update.md. Four lines: what I did, what I found, what is blocking me, what I am doing next. This takes ten minutes and pays off repeatedly — meetings start from a shared artifact instead of from your recollection, you build a dated record of the project that makes writing much easier, and problems surface while they are still small.

Budget two to three weeks of turnaround per chapter, and plan around it. This is the concrete scheduling implication: send chapters as each finishes rather than batching them, because four chapters at two weeks each is eight weeks of serial latency if you send them together and much less if you pipeline. When you send something, say what kind of feedback you want and by when.

Never go dark. If a month has gone badly and you have nothing to show, send the update anyway saying so. Silence is read as trouble and it removes the one person positioned to help you. This is the most common avoidable failure in graduate work.

Ask directly about the things that are actually ambiguous: what your committee expects, what "done" looks like for your program, and whether your current scope is sufficient. Do not infer these. Ask, and ask early enough that the answer can change what you do.

When feedback seems wrong, engage rather than either capitulating or ignoring it. Ask what concern is behind the comment. Often a suggestion you disagree with points at a real problem that a different fix would solve better.

Tooling

LaTeX, with the skeleton in thesis/. Use your department's template if one exists.

One sentence per line in your source files. This is not a style preference — it makes git diffs readable at sentence granularity, which makes version control on prose genuinely useful. Reflowed paragraphs produce diffs you cannot review.

Commit daily with real messages. The point is being able to recover a paragraph you deleted in week thirty-four and want back in week thirty-eight.

latexdiff produces a marked-up PDF between two versions. Send this to your advisor for revision rounds so they can see what changed instead of rereading the chapter.

Figures as vector PDF, generated by scripts, with fonts matching your document size so labels are readable in print. Never screenshot a plot.

Reference management as described in guide/02-reading-system.md: Zotero plus Better BibTeX auto-exporting to thesis/refs.bib.

Academic integrity

Two points that matter more than they used to.

Paraphrasing is not rewording. Changing a few words in someone's sentence is plagiarism even with a citation attached. Read, close the paper, write the idea in your own words from understanding, then cite. If you want their exact phrasing, quote it explicitly with a page number. Self-plagiarism is also a real category: if you reuse text from your own published paper, check your institution's rules, since many require disclosure.

Find and follow your institution's policy on AI writing tools. These policies exist now, they vary substantially, and they are frequently specific about disclosure. Read yours rather than assuming. Independent of policy, be aware of two failure modes: generated text tends to be fluent and vague in ways that read as padding to an expert, and generated citations are sometimes fabricated — verify that every reference in your bibliography exists and says what you claim it says.

Whatever your policy permits, every claim in your thesis must be one you can defend under questioning, because you will be questioned. That constraint is the practical one.