paper-writing
Draft a CV/graphics conference paper section by section using outline-first writing — list the thinking down to sentence granularity in the user's own language first, then produce English prose. Ships LaTeX skeleton templates with real published examples (PVNet, Deep Snake, Neural Body, ManhattanSDF, DreamBooth) for abstract, introduction, technical-challenge framing, method, experiments, related work, conclusion and title, plus a fully worked four-granularity outline and the 11-step order the sections get written in. Use whenever the user is writing or planning any part of a paper, asks how to introduce a technical challenge or motivate a module, wants help with the paper's story, or asks which section to write next — and reach for it before writing any English paragraph, since drafting prose before the outline is the single most common cause of unfixable structure. Boundaries: deadline tracking and the submission schedule belong to research-project; if the text already exists and the complaint is that it re
소스 정보
- 저장소
- HughYau/pengsida-learning-research-skills
- 최근 소스 활동
- 2026년 8월 9일 08:48
- 감지된 SKILL.md 언어
- 중국어
- 스타
- 14
- 포크
- 1
설치 방법
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
소스 파일 검토
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
파일 탐색기
9 개 파일SKILL.md 표시 중
SKILL.md
소스 지침 · 읽기 전용 미리보기- name
- paper-writing
- description
- Draft a CV/graphics conference paper section by section using outline-first writing — list the thinking down to sentence granularity in the user's own language first, then produce English prose. Ships LaTeX skeleton templates with real published examples (PVNet, Deep Snake, Neural Body, ManhattanSDF, DreamBooth) for abstract, introduction, technical-challenge framing, method, experiments, related work, conclusion and title, plus a fully worked four-granularity outline and the 11-step order the sections get written in. Use whenever the user is writing or planning any part of a paper, asks how to introduce a technical challenge or motivate a module, wants help with the paper's story, or asks which section to write next — and reach for it before writing any English paragraph, since drafting prose before the outline is the single most common cause of unfixable structure. Boundaries: deadline tracking and the submission schedule belong to research-project; if the text already exists and the complaint is that it reads badly, that is paper-revision. 中文触发：写论文、写introduction、写abstract、写method、写experiment、写related work、梳理story、写作思路、论文标题、先写哪个章节、开始写了。
# 论文写作
**核心方法：先列清楚写作思路，理清要表达的内容，再写英语句子。**
写作思路要细化到知道每一句话写什么内容。
不要打开编辑器就开始写英文——那样写出来的段落逻辑必然是拼凑的，
后期改起来成本极高。
## 写作顺序（11 步）
| # | 步骤 | 参考 |
|---|---|---|
| 1 | 画一个清楚的 pipeline figure 草图 | [references/figures-and-tables.md](references/figures-and-tables.md) |
| 2 | 梳理论文 story，写 introduction 的写作思路；整理要做的 comparison / ablation | [references/introduction.md](references/introduction.md) |
| 3 | 列 method 写作思路 → 写 method（同时做实验） | [references/method.md](references/method.md) |
| 4 | 改 introduction 和 method（同时做实验） | `paper-revision` |
| 5 | 实验做得差不多后，列 experiment 思路 → 写 experiment | [references/experiments.md](references/experiments.md) |
| 6 | 美化 pipeline figure，画 teaser 图 | [references/figures-and-tables.md](references/figures-and-tables.md) |
| 7 | 列 related work 思路 → 写 related work | [references/related-work-conclusion-title.md](references/related-work-conclusion-title.md) |
| 8 | Review 论文，改 introduction / method / experiment | `paper-self-review` |
| 9 | 列 abstract 思路 → 写 abstract | [references/abstract.md](references/abstract.md) |
| 10 | 取论文标题 | [references/related-work-conclusion-title.md](references/related-work-conclusion-title.md) |
| 11 | 反复 review、反复改 | `paper-self-review` + `paper-revision` |
注意 abstract 排在很后面（第 9 步）——它是全文的压缩，story 没定之前写不出来。
顺序能推进的前提是**不被未定的细节卡住**：method 里还没定下来的地方写 `\todo{}`
空着，先把框架写出来。框架成形了才能判断 story 通不通，而细节后补的成本很低；
反过来，为了一个没定的细节停下不写，整篇的结构问题就要拖到最后才暴露，
那时候改最贵。（`research-project` 的 `project_status.py` 正是数各 section 的字数
和 `\todo{}` 数量来估章节进度的——空着但写了 `\todo{}` 和真的没写，它分得出来。）
## 写作思路的四个粒度
从粗到细依次列出，每一层稳定了再进入下一层：
1. **Section 粒度**：每个 section 写什么。（比较确定：Abstract / Introduction /
Related work / Method / Experiment）
2. **Sub-section 粒度**：每个 sub-section 写什么。（一般只有 Method 和 Experiment 有）
3. **段落粒度**：每个段落写什么。每个段落只能有**一个**要表达的信息，段落间逻辑要流畅。
4. **句子粒度**：每句话写什么。句子间逻辑要流畅，内容要完整。
完整的四粒度实例（ManhattanSDF 的 Introduction / Related work / Method 全文思路）：
[references/writing-outline-example.md](references/writing-outline-example.md)。
**这是本 skill 最有价值的单个文件**，需要产出写作思路时优先参考它的形态。
## 段落写作原则
1. **一段文字只讲一个 Message**，并表达清楚，不要把几个 Messages 杂糅在一起。
2. **段落开头第一句就要让读者知道这段在说什么。**
（金字塔原理：塔尖是想传达的观点，塔基是支撑观点的逻辑论据）
3. 注意段落之间、句子之间的 **flow**。
## 用 AI 辅助英语写作的正确顺序
这是 LLM 时代的基本技能，但顺序不能反：
```
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
```
第 e 步是最容易被跳过、也最关键的一步：写具体句子时才会暴露思路的漏洞，
此时要回头改思路，而不是硬把句子写通。实际写作中一个段落的思路迭代 4-5 版是正常的。
## 两条贯穿全文的要求
> **论文收获好 review 的关键：把论文做得漂亮、美观，让人第一印象觉得这篇论文很高级。**
> 三个抓手：好看的 teaser figure / pipeline figure；好看的表格和结果图；整齐的排版。
> **论文中所有的 claim（特别是 abstract 和 introduction 里的）必须不犯错，且有实验 support。**
> 否则一些 reviewer 会直接以此拒掉论文。
**写论文一定要"如切如磋，如琢如磨"，反复品味，揣摩读者是否读得懂。**
"自我评阅论文写作是否清楚"的能力非常重要——知道有问题才能知道要改。
判断写得清不清楚、以及怎么改，见 `paper-revision`。
## 与其他 skill 的衔接
| 阶段 | 用哪个 |
|---|---|
| 写作思路与章节模板 | 本 skill 的 `references/` |
| 改句子、改思路、判断 flow | `paper-revision` |
| 投稿前自评审 / 回应审稿 | `paper-self-review` / `paper-revision` |
| 写作产物落到项目文件 | `research-project`：`docs/writing/outline.md`、`PROJECT.md` §5 |
| **单张图的正确性与可读性** | `figure-style`——先 `apply_figure_style()` 再画 |
| **多子图编排与自审** | `figure-composer` |
| **整套图讲的故事对不对** | `paper-narrative`——以 handling editor 视角审 arc |
| 补 related work 的检索 | `literature-review` |
| 做学术报告 slides | [references/talk-slides.md](references/talk-slides.md)（同构于 Introduction） |
关于图的分工：本 skill 的 [references/figures-and-tables.md](references/figures-and-tables.md)
给的是 **CV/graphics 投稿的修辞要求**（pipeline 图要和前人不一样、要突出 novel module、
表格用 booktabs）；`figure-style` 给的是**技术正确性与可读性**（数据保真、色觉障碍、
字号阶梯）。两者管的是不同的问题，都要满足；真冲突时按 `research-methodology`
的裁决规则走——数据保真和色觉障碍友好没有商量余地。
GitHub에서 보기