write

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Write

写作

Draft the paper from

RESEARCH.md

. You can start writing at any point in the research lifecycle; use

[UNGROUNDED]

to track claims that still need evidence.

从

RESEARCH.md

起草论文。你可以在研究周期的任意阶段开始写作；使用

[UNGROUNDED]

标记仍需补充证据的论点。

Prerequisites

前置条件

Read
```
RESEARCH.md
```
. Apply supervision policy for
```
write-start
```
(approve → pause; notify → announce and continue; absent → proceed). If writing requires a compromise relative to the current plan, apply
```
Idea Changes
```
before proceeding.
Check Context for experiment results.
- Results exist → use as empirical basis.
- No results → ask: proceed with a results-free draft, or wait for experiments?

阅读
```
RESEARCH.md
```
。执行
```
write-start
```
监督策略（批准→暂停；通知→告知并继续；无指示→继续）。如果写作需要偏离当前计划，需先执行
```
Idea Changes
```
流程再继续。
查看Context中的实验结果。
- 已有结果→作为实证依据使用。
- 无结果→询问：是继续撰写无结果的草稿，还是等待实验完成？

Narrative Framing

叙事框架

Before outlining, state the paper's contribution in one sentence. If you cannot, the framing has not converged — ask the user to clarify the intended story.
Build a compact claim-evidence map:
- Claim — what the paper asserts.
- Evidence — result, theorem, artifact, or citation supporting it.
- Section — where the claim will appear.
- Missing evidence → mark
```
[UNGROUNDED]
```
  and keep drafting only if the user approved a results-free draft.

在拟定大纲前，用一句话阐明论文的贡献。如果无法做到，说明叙事框架尚未明确——请用户澄清预期的核心故事。
构建简洁的论点-证据映射表：
- 论点——论文主张的内容。
- 证据——支持论点的实验结果、定理、成果或引用文献。
- 章节——论点将出现的位置。
- 缺失证据→标记
```
[UNGROUNDED]
```
  ，仅当用户批准撰写无结果草稿时才可继续起草。

Drafting Order

起草顺序

Follow this sequence strictly:
1. Outline — section structure with bullet points for each section's argument. The outline should make one coherent story, not a list of experiments.
2. Section drafts — write each section, starting from the one with the strongest evidence.
3. Self-consistency check — verify that claims in the abstract match claims in the body, numbers are consistent across sections, and the conclusion follows from the results.
4. Citation pass — see below.

严格遵循以下步骤：
1. 大纲——列出章节结构，每个章节的论点用项目符号表示。大纲应呈现连贯的叙事逻辑，而非实验列表。
2. 章节草稿——撰写每个章节，从证据最充分的章节开始。
3. 自洽性检查——验证摘要中的论点与正文一致，各章节中的数据统一，结论由结果推导而来。
4. 引用检查——详见下文。

Grounding Rule

依据规则

Every quantitative claim, dataset claim, baseline claim, theorem claim, and method-detail claim must trace to a specific source in
```
RESEARCH.md
```
Context or a verified citation. If a claim cannot be traced, mark it inline:
```
[UNGROUNDED]
```
. Leave the marker visible for the user — do not silently remove it.

每一个量化论点、数据集论点、基准论点、定理论点和方法细节论点都必须能追溯到
```
RESEARCH.md
```
的Context或已验证的引用来源。如果论点无法追溯，在文中标记：
```
[UNGROUNDED]
```
。保留该标记供用户查看——不得擅自删除。

Style Pass

风格优化

After drafting, polish for publication-ready academic prose:
- Keep subject and verb close.
- Use concrete verbs and consistent terminology.
- Replace vague nouns like "performance" with the actual metric when known.
- Remove filler and inflated language: "groundbreaking", "revolutionary", "delve", "pivotal", "landscape", "notably", "importantly", "it is worth noting".
- Avoid generic conclusions such as "opens new avenues" unless the avenues are specific.
- Run a reverse-outline test: read the first sentence of every paragraph in sequence. They should form a coherent argument on their own.

起草完成后，润色为可发表的学术文风：
- 主语与谓语尽量靠近。
- 使用具体动词和统一术语。
- 已知具体指标时，替换“性能”等模糊名词。
- 删除冗余和夸大表述：如“突破性的”“革命性的”“深入研究”“关键的”“格局”“值得注意的是”等。
- 避免泛泛的结论，如“开辟新方向”，除非明确说明具体方向。
- 执行反向大纲测试：依次阅读每个段落的第一句，它们应能独立构成连贯的论证。

Citation Pass

引用检查

For every reference: search arXiv, resolve the DOI, or use another trusted source. If unverifiable, mark
```
[UNVERIFIED]
```
. Never generate citations from memory. Fabricated citations are academic misconduct.

对于每一处引用：搜索arXiv、解析DOI或使用其他可信来源。如果无法验证，标记
```
[UNVERIFIED]
```
。切勿凭记忆生成引用。伪造引用属于学术不端行为。

Output

输出

Default output:
```
paper.md
```
and
```
refs.bib
```
at the project root. Escape hatch: if a
```
.tex
```
file already exists at the project root, ask whether to write there instead.
Before declaring done, check:
- Abstract is self-contained.
- Introduction claims have supporting evidence in the body.
- Numbers are consistent across sections.
- Terminology is consistent.
- ```
[UNGROUNDED]
```
  and
```
[UNVERIFIED]
```
  markers are listed for the user.

默认输出：项目根目录下的
```
paper.md
```
和
```
refs.bib
```
文件。 应急方案：如果项目根目录已存在
```
.tex
```
文件，询问是否改为写入该文件。
在完成前检查：
- 摘要独立完整。
- 引言中的论点在正文有支撑证据。
- 各章节数据统一。
- 术语统一。
- 已向用户列出
```
[UNGROUNDED]
```
  和
```
[UNVERIFIED]
```
  标记。

Post-Draft

草稿后处理

Auto-invoke
```
review
```
. Write the score to
```
RESEARCH.md
```
Context. Apply stop policy (
```
target_reached
```
). In wild mode, continue if configured; otherwise ask: "Review score is X/10. Want a revision pass?"

自动调用
```
review
```
功能。将评分写入
```
RESEARCH.md
```
的Context。执行停止策略（
```
target_reached
```
）。在wild模式下，若已配置则继续；否则询问：“评分为X/10。是否需要修订？”

Partial State

部分状态处理

If
```
paper.md
```
already exists, ask whether to continue from it or start fresh.

如果
```
paper.md
```
已存在，询问是继续基于该文件撰写还是重新开始。

Constraints

约束条件

Never fabricate results, citations, baselines, or venue claims.
Never hide
```
[UNGROUNDED]
```
or
```
[UNVERIFIED]
```
markers.
Do not turn
```
write
```
into compilation, figure generation, or experiment execution; delegate those to the relevant skill.

切勿伪造结果、引用、基准或期刊/会议相关声明。
不得隐藏
```
[UNGROUNDED]
```
或
```
[UNVERIFIED]
```
标记。
不得将
```
write
```
功能用于编译、图表生成或实验执行；这些任务需委托给相关技能。

Example

示例

Input: RESEARCH.md with 2 completed experiments; user says "write the paper". Output: paper.md (publication-ready narrative + 5 sections, 1 [UNGROUNDED] marker), refs.bib (8 entries, 1 [UNVERIFIED]), then auto-invokes

review

输入：包含2项已完成实验的RESEARCH.md；用户指令“write the paper”。输出：paper.md（可发表叙事结构+5个章节，含1个[UNGROUNDED]标记）、refs.bib（8条条目，含1个[UNVERIFIED]标记），随后自动调用

review

。