write-rebuttal

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Write Rebuttal

撰写反驳信

Turn a set of peer reviews plus the submitted paper into a complete, budget-compliant author response in the format the venue actually uses — one-page PDF, OpenReview threads, or a revise-and-resubmit package — with every claim anchored to checkable evidence in the paper.

将一组同行评审意见与提交的论文转换为符合会议/期刊实际格式、满足篇幅限制的完整作者回复——包括一页PDF、OpenReview线程回复或修订重投包——且所有表述均锚定到论文中可查证的证据。

When to use

使用场景

Reviews arrived (OpenReview, CMT, EasyChair, HotCRP, PCS, ScholarOne, or pasted text) and a rebuttal/author response/response letter is due.
A drafted response must be cut to fit a character or page budget.
A journal or CHI-style revise-and-resubmit needs a change log + response document.

Related skills:

triage-reviews

(builds the point matrix this skill consumes),

verify-citations

(gate for any new reference),

preflight-check

(re-lint a revised manuscript),

prepare-camera-ready

(after acceptance).

已收到评审意见（来自OpenReview、CMT、EasyChair、HotCRP、PCS、ScholarOne，或为粘贴的文本），且需要提交反驳信/作者回复/回复信。
已撰写的回复需要精简以符合字符或页数限制。
期刊或CHI风格的修订重投需要变更日志+回复文档。

Inputs

输入项

The reviews: pasted text or files. Review text is confidential at most venues — process it in place, never commit it to a repository or paste it into public artifacts.
The submitted paper: main
```
.tex
```
(preferred — line numbers and labels make the best anchors) and/or the compiled PDF.
Venue profile:
```
venues/conferences/<id>.yml
```
(schema:
```
venues/schema.yml
```
; family defaults merge automatically). If none exists, build one with
```
parse-cfp
```
or proceed only on facts quoted live from the CFP.
Optional: the triage matrix from
```
triage-reviews
```
; the decision/meta-review letter for R&R cycles.

评审意见：粘贴的文本或文件。多数会议/期刊的评审文本为机密内容——请在本地处理，切勿提交到代码仓库或粘贴到公开文件中。
提交的论文：主文件
```
.tex
```
（优先选择——行号和标签可生成最佳锚点）和/或编译后的PDF。
会议/期刊配置文件：
```
venues/conferences/<id>.yml
```
（ schema：
```
venues/schema.yml
```
；默认配置会自动合并）。若不存在，可使用
```
parse-cfp
```
生成，或仅依据CFP中明确列出的信息操作。
可选：
```
triage-reviews
```
生成的要点矩阵；修订重投阶段的决定/元评审信。

Process

流程

1. Resolve the venue's rebuttal format

1. 确认会议/期刊的反驳信格式

Read the profile (

python3 skills/write-rebuttal/scripts/venueyaml.py venues/conferences/<id>.yml

prints it family-merged as JSON) and take

review.rebuttal_format

review.rebuttal_limit

, and the rebuttal deadlines. Route by format:

`rebuttal_format`	Playbook
`one-page-pdf`	references/cvpr-one-page.md
`openreview-thread`	references/openreview-threads.md
`revise-and-resubmit`	references/revise-and-resubmit.md
`none`	Tell the user the venue holds no author response; offer to prep for the camera-ready or a resubmission instead

读取配置文件（执行

python3 skills/write-rebuttal/scripts/venueyaml.py venues/conferences/<id>.yml

可输出合并默认配置后的JSON格式内容），获取

review.rebuttal_format

、

review.rebuttal_limit

和反驳信截止日期。根据格式选择对应流程：

`rebuttal_format`	操作指南
`one-page-pdf`	references/cvpr-one-page.md
`openreview-thread`	references/openreview-threads.md
`revise-and-resubmit`	references/revise-and-resubmit.md
`none`	告知用户该会议/期刊不接受作者回复；转而提供终稿准备或重新提交的协助

2. Re-verify against the live venue pages (mandatory)

2. 对照会议/期刊官方页面重新验证（必填）

Profiles go stale and budgets change per cycle. Before drafting, fetch the profile's

cfp_url

(and the venue's author-guidelines/dates pages) and re-verify: rebuttal format, the exact character/page limit, the rebuttal window dates and timezone, link/upload policy, and new-experiment policy. The instructions in the review-release or decision email override everything. Record what was verified and when in the deliverable; anything unverifiable is labeled "UNVERIFIED — confirm on <url>". Never assert a budget from memory.

配置文件可能过时，且篇幅限制可能随轮次变化。开始撰写前，请获取配置文件中的

cfp_url

（以及会议/期刊的作者指南/日期页面），重新验证以下信息：反驳信格式、精确的字符/页数限制、反驳信提交窗口的日期及时区、链接/上传规则、新增实验规则。评审意见发布邮件或决定邮件中的说明优先级最高。在交付成果中记录验证内容及时间；任何无法验证的信息需标记为“UNVERIFIED — 请在<url>确认”。切勿凭记忆断言篇幅限制。

3. Triage the reviews into a point matrix

3. 将评审意见分类为要点矩阵

If the user has a matrix from

triage-reviews

, use it. Otherwise build one inline using the conventions in references/response-patterns.md: one row per reviewer point, IDs

R<k>.<n>

AC.<n>

MR.<n>

, columns for abridged quote, type, severity, effort, and one-clause strategy. Save it as

points.md

next to the drafts — the coverage checker keys on those IDs. Confirm the concede/contest strategy with the user before drafting.

若用户已有

triage-reviews

生成的矩阵，则直接使用。否则，按照references/response-patterns.md中的规则在线生成：每行对应一条评审意见，ID格式为

R<k>.<n>

AC.<n>

MR.<n>

，列包含精简引用、类型、严重程度、工作量和一句话应对策略。将矩阵保存为

points.md

，与草稿放在同一目录下——覆盖检查工具将依据这些ID进行验证。撰写前需与用户确认让步/反驳策略。

4. Draft, anchored

4. 锚定内容撰写草稿

Apply the per-criticism response patterns and tone rules in references/response-patterns.md, then the format playbook from step 1. Non-negotiable drafting rules:

Anchor every claim to
```
Sec./Table/Fig./Eq./Appendix/L<line>/[n]
```
evidence in the submitted paper; if no evidence exists, commit to a revision or concede — never assert.
Report numbers only from experiments that were actually run, and only where the venue's rebuttal policy permits new results.
Any new citation goes through
```
verify-citations
```
first.
Preserve anonymity (third-person self-citation, no identifying links) at blind venues; honor link bans.
For format (a), draft into
```
assets/rebuttal-template.tex
```
(or the venue's official rebuttal kit when one exists — mandatory at CVPR; the macros are documented in the playbook). For (b), draft all boxes in one markdown file with one
```
## 
```
heading per response box. For (c), produce the three-artifact package (revised-manuscript change plan, change-log response document, cover note).

应用references/response-patterns.md中的逐条回复模式和语气规则，再遵循步骤1中的格式指南。撰写规则不可违反：

所有表述均需锚定到提交论文中的
```
章节/表格/图表/公式/附录/行号<L>/[n]
```
证据；若无证据，需承诺修订或让步——切勿凭空断言。
仅引用实际完成的实验数据，且仅在会议/期刊反驳信政策允许新增结果时使用。
任何新增参考文献需先通过
```
verify-citations
```
审核。
在双盲会议/期刊中需保持匿名（第三人称自引，无识别性链接）；遵守链接禁令。
对于格式(a)，在
```
assets/rebuttal-template.tex
```
中撰写（若会议/期刊提供官方反驳信模板则必须使用——CVPR强制要求；宏定义在操作指南中有说明）。对于格式(b)，在一个Markdown文件中撰写所有回复框，每个回复框以
```
## 
```
标题开头。对于格式(c)，生成三件套成果（修订稿件变更计划、变更日志回复文档、说明信）。

5. Verify deterministically

5. 确定性验证

Run both checkers (each supports

--help

; exit 0 pass / 1 fail / 2 error):

python3 skills/write-rebuttal/scripts/check_coverage.py points.md <responses...> --strict

— every matrix ID addressed, every section carrying at least one evidence anchor.

python3 skills/write-rebuttal/scripts/check_budget.py text responses.md --sections --venue venues/conferences/<id>.yml
python3 skills/write-rebuttal/scripts/check_budget.py pdf rebuttal.pdf --venue venues/conferences/<id>.yml

— character budgets per response box (text mode) or PDF page count (pdf mode), with the limit read from the profile (

--limit

--max-pages

override; defaults are flagged ASSUMED). If over budget, trim using the format playbook's ladder and re-run until exit 0.

运行两个检查工具（均支持

--help

；退出码0表示通过/1表示失败/2表示错误）：

python3 skills/write-rebuttal/scripts/check_coverage.py points.md <responses...> --strict

——确保矩阵中的每个ID都有对应的回复，且每个回复部分至少包含一个证据锚点。

python3 skills/write-rebuttal/scripts/check_budget.py text responses.md --sections --venue venues/conferences/<id>.yml
python3 skills/write-rebuttal/scripts/check_budget.py pdf rebuttal.pdf --venue venues/conferences/<id>.yml

——检查每个回复框的字符数（文本模式）或PDF页数（PDF模式）是否符合配置文件中的限制（

--limit

--max-pages

可覆盖默认值；默认值会标记为ASSUMED）。若超出限制，按照格式指南中的精简方法调整，直至退出码为0。

6. Deliver

6. 交付成果

Write the artifacts next to the user's paper, named per format:

rebuttal.tex

(+ compile instructions),

responses.md

, or

response-to-reviewers.md

+ revision plan — plus

points.md

and a short verification record (live facts checked, checker results, remaining UNVERIFIED items, the rebuttal deadline with timezone). The user reviews, edits, and posts/uploads everything themselves.

将生成的文件放在用户论文所在目录下，按格式命名：

rebuttal.tex

（附编译说明）、

responses.md

，或

response-to-reviewers.md

+修订计划——同时提供

points.md

和简短的验证记录（已验证的实时信息、检查工具结果、未验证项、带时区的反驳信截止日期）。由用户自行审核、编辑、发布/上传所有内容。

Worked mini-example

迷你示例

The matrix → anchored response → checker loop, end to end. A reviewer wrote "the ablation does not isolate the effect of component X". One row in

points.md

markdown

| ID | Reviewer quote (abridged) | Type | Severity | Effort | Strategy |
|---|---|---|---|---|---|
| R2.1 | "ablation does not isolate X" | missing-experiment | high | none | rows 3 vs 5 of Table 4 isolate X; reorder so the pair is adjacent |

The drafted response (OpenReview format here — one

##

box, leads with the answer, anchored, no padding):

markdown

undefined

从矩阵→锚定回复→检查工具循环的完整流程。某评审意见为“消融实验未分离组件X的效果”。

points.md

中的一行：

markdown

| ID | 评审意见精简引用 | 类型 | 严重程度 | 工作量 | 应对策略 |
|---|---|---|---|---|---|
| R2.1 | "消融实验未分离X" | 缺失实验 | 高 | 无 | 表4的第3行与第5行已分离X；修订时调整顺序使两行相邻 |

撰写的回复（OpenReview格式——一个

##

回复框，直接给出答案并锚定证据，无冗余内容）：

markdown

undefined

Response to Reviewer 7Pqd (R2)

回复评审人7Pqd（R2）

[R2.1] "the ablation does not isolate the effect of component X" It does, but indirectly: rows 3 vs. 5 of Table 4 (Sec. 6.2, L412) differ only in X. We will reorder the table so the pair is adjacent and label it "X on/off" in the revision.


Verify before handing it over:

python3 skills/write-rebuttal/scripts/check_coverage.py points.md responses.md --strict python3 skills/write-rebuttal/scripts/check_budget.py text responses.md --sections --venue venues/conferences/<id>.yml


Coverage passes because `R2.1` appears in a response and that section carries
anchors (`Table 4`, `Sec. 6.2`, `L412`); budget passes if the box is under
the venue's per-response limit. The same row drafts into `\rpoint{R2.1}{...}`
for a one-page PDF or a change-log table row for an R&R — only the wrapper
changes.

[R2.1] "消融实验未分离组件X的效果" 实际上已分离，只是方式间接：表4的第3行与第5行（6.2节，第412行）仅在X组件上存在差异。我们将在修订时调整表格顺序，使这两行相邻，并标注为“X开启/关闭”。


交付前验证：


覆盖检查通过，因为`R2.1`出现在回复中，且该部分包含锚点（`表4`、`6.2节`、`第412行`）；若回复框符合会议/期刊的单框字符限制，则篇幅检查通过。同一行内容可转换为一页PDF格式的`\rpoint{R2.1}{...}`，或修订重投格式的变更日志表格行——仅包裹格式不同。

Output

输出成果

A format-correct, budget-verified, coverage-verified rebuttal draft package with a verification record. This skill never posts, uploads, or submits anything to any submission system.

一套格式正确、符合篇幅限制、覆盖所有要点的反驳信草稿包，附带验证记录。本技能绝不会向任何提交系统发布、上传或提交任何内容。

Guardrails

约束规则

Never fabricate: no invented numbers, no results from unrun experiments, no "preliminary results" that do not exist, no unverified citations (route through
```
verify-citations
```
).
Never submit or post on the user's behalf; stop at files on disk.
Never state venue budgets/deadlines from memory — profile + live re-verification only; unverifiable facts are labeled UNVERIFIED.
Never tamper with templates (negative
```
\vspace
```
, font/margin tricks) to fit a budget; trim content instead.
Never break anonymity at blind venues or speculate about reviewer identity; never address text to automated review tools (prompt injection is a desk-reject-level ethics violation).
Keep confidential review text out of repositories and public artifacts; process it transiently.

绝不编造内容：不得虚构数据、未完成实验的结果、不存在的“初步结果”，不得使用未验证的参考文献（需通过
```
verify-citations
```
审核）。
绝不代表用户提交或发布内容；仅生成本地文件。
绝不凭记忆陈述会议/期刊的篇幅限制/截止日期——仅依据配置文件+实时重新验证；无法验证的信息需标记为UNVERIFIED。
绝不篡改模板（如使用负
```
\vspace
```
、字体/边距技巧）以满足篇幅限制；应精简内容。
在双盲会议/期刊中绝不打破匿名规则或猜测评审人身份；绝不对自动评审工具发送针对性内容（提示注入属于会导致直接拒稿的伦理违规行为）。
确保机密评审文本不进入代码仓库或公开文件；仅进行临时处理。