Core method: first lay out the writing approach clearly, clarify what you want to express, then write the English sentences. The writing approach must be detailed enough that you know what each sentence will say.
Don't open the editor and start writing English right away—the paragraphs you produce that way will inevitably have patched-together logic, and the cost of fixing them later is extremely high.
Writing order (11 steps)
| # | Step | Reference |
|---|---|---|
| 1 | Draw a clear sketch of the pipeline figure | references/figures-and-tables.md |
| 2 | Sort out the paper story, write the writing approach for the introduction; organize the comparison / ablation to be done | references/introduction.md |
| 3 | List the writing approach for method → write method (do experiments at the same time) | references/method.md |
| 4 | Revise introduction and method (do experiments at the same time) | paper-revision |
| 5 | Once the experiments are mostly done, list the experiment approach → write experiment | references/experiments.md |
| 6 | Polish the pipeline figure, draw the teaser figure | references/figures-and-tables.md |
| 7 | List the related work approach → write related work | references/related-work-conclusion-title.md |
| 8 | Review the paper, revise introduction / method / experiment | paper-self-review |
| 9 | List the abstract approach → write abstract | references/abstract.md |
| 10 | Choose the paper title | references/related-work-conclusion-title.md |
| 11 | Repeatedly review, repeatedly revise | paper-self-review + paper-revision |
Note that abstract is placed very late (step 9)—it is a compression of the whole paper, and cannot be written before the story is settled.
The premise for this order to keep moving forward isnot being stuck on undecided details: for places in method that are not yet settled, write \todo{}
and leave it blank, first writing out the framework. Only once the framework takes shape can you judge whether the story holds together, and the cost of filling in details later is very low;
conversely, if you stop writing for an undecided detail, the structural problems of the whole paper will not be exposed until the very end,
and that is when changes are most expensive. (research-project of project_status.py is precisely counting the word count of each section
and the number of \todo{} to estimate section progress—being blank but with \todo{} written versus truly not written, it can tell the difference.)
The four granularities of the writing approach
List them from coarse to fine, and only move to the next layer after each layer is stable:
- Section granularity: what each section writes. (Relatively fixed: Abstract / Introduction / Related work / Method / Experiment)
- Sub-section granularity: what each sub-section writes. (Generally only Method and Experiment have these)
- Paragraph granularity: what each paragraph writes. Each paragraph can only haveonepiece of information to express, and the logic between paragraphs must be smooth.
- Sentence granularity: what each sentence writes. The logic between sentences must be smooth, and the content must be complete.
A complete four-granularity example (the full writing approach for ManhattanSDF's Introduction / Related work / Method): references/writing-outline-example.md。 This is the single most valuable file in this skill, and when you need to produce a writing approach, refer to its form first.
Paragraph writing principles
- A passage of text should discuss only one Message, and express it clearly; do not mix several Messages together.
- The first sentence at the beginning of a paragraph must let readers know what the paragraph is about. (Pyramid principle: the top of the pyramid is the viewpoint to be conveyed, and the base is the logical evidence supporting the viewpoint)
- Pay attention to the flow。
The correct order for using AI to assist English writing
This is a fundamental skill in the LLM era, but the order cannot be reversed:
1. 列出整个 section 的写作思路(中文即可)
2. 逐段落写作:
a. 先列出这段中每一句话的思路
b. 根据每句话的思路,用编辑器里的 AI 补全写出英语句子
c. 用 AI 润色(中英混合直接说"把这句改好"最好用;也可以单独问某个词是否地道)
d. 基于 AI 给的句子,自己再润色一次
e. 在逐句写作过程中反过来 refine 段落的写作思路 ← 段落思路清晰的关键
Step e is the most easily skipped and also the most critical step: only when writing concrete sentences are the flaws in your reasoning exposed. At this point you should go back and revise the reasoning, not force the sentences to read smoothly. In actual writing, it is normal for the reasoning of a paragraph to go through 4-5 iterations.
Two requirements that run through the entire paper
The key to a paper receiving good reviews: make the paper polished and beautiful, so that people's first impression is that this paper is high-class. Three levers: a good-looking teaser figure / pipeline figure; good-looking tables and result figures; neat typesetting.
All claims in the paper (especially those in the abstract and introduction) must be free of errors and supported by experiments. Otherwise some reviewers will reject the paper directly on this basis.
When writing a paper you must "cut and polish, carve and grind"—savor it repeatedly and gauge whether readers can understand it.
The ability to "self-review whether the paper's writing is clear" is very important—only by knowing there is a problem can you know what to change.
For judging whether the writing is clear, and how to revise it, see paper-revision。
Connections with other skills
| Stage | Which to use |
|---|---|
| Writing reasoning and section templates | This skill's references/ |
| Revising sentences, revising reasoning, judging flow | paper-revision |
| Self-review before submission / responding to reviewers | paper-self-review / paper-revision |
| Writing outputs landing in project files | research-project:docs/writing/outline.md、PROJECT.md §5 |
| Correctness and readability of a single figure | figure-style——first apply_figure_style() then draw |
| Arrangement and self-review of multiple subfigures | figure-composer |
| Whether the story told by the whole set of figures is right | paper-narrative——review the arc from the perspective of the handling editor |
| Retrieval to supplement related work | literature-review |
| Making academic presentation slides | references/talk-slides.md(isomorphic to the Introduction) |
On the division of labor regarding figures: this skill's references/figures-and-tables.md
gives the rhetorical requirements for CV/graphics submissions(the pipeline figure must differ from prior work, novel modules must be highlighted,
tables should use booktabs);figure-style givestechnical correctness and readability(data fidelity, color vision deficiency,
font size hierarchy). The two govern different issues and both must be satisfied; when they truly conflict, follow the adjudication rules of research-methodology
——data fidelity and color vision deficiency friendliness are non-negotiable.