citation-audit

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Citation Audit

引用审核

Run a pre-submission audit of citations, BibTeX entries, and LaTeX cross-references. This skill is for checking correctness before submission, not for broad literature discovery.
Use this skill when a paper already has draft citations and the user wants confidence that:
  • every
    \cite{...}
    key in TeX exists in BibTeX
  • every BibTeX entry is syntactically valid and not duplicated
  • every
    \ref{...}
    ,
    \cref{...}
    ,
    \eqref{...}
    ,
    \autoref{...}
    target exists
  • every
    \label{...}
    is unique and follows the local naming convention
  • DOI, arXiv, OpenReview, URL, title, author, year, and venue metadata match the real paper
  • citation claims in nearby prose are actually supported by the cited work
  • bibliography style is submission-ready
Pair this with
submit-paper
for the broader submission checklist. Pair it with
research-project-memory
when citation correctness issues should become blocking paper risks or actions.
为LaTeX学术论文的引用、BibTeX条目以及交叉引用执行提交前审核。此技能用于提交前检查正确性,而非广泛的文献发现。
当论文已有 draft 引用,且用户希望确认以下内容时使用此技能:
  • TeX中的每个
    \cite{...}
    键在BibTeX中都存在
  • 每个BibTeX条目语法有效且无重复
  • 每个
    \ref{...}
    \cref{...}
    \eqref{...}
    \autoref{...}
    的目标都存在
  • 每个
    \label{...}
    唯一且遵循本地命名规范
  • DOI、arXiv、OpenReview、URL、标题、作者、年份和会议/期刊元数据与实际论文匹配
  • 附近文本中的引用声明确实得到被引用作品的支持
  • 参考文献格式符合提交要求
可与
submit-paper
技能配合使用,完成更全面的提交检查清单。当引用正确性问题应被列为论文的阻塞性风险或待办事项时,可与
research-project-memory
技能配合使用。

Skill Directory Layout

技能目录结构

text
<installed-skill-dir>/
├── SKILL.md
├── scripts/
│   └── audit_latex_refs.py
└── references/
    ├── citation-claim-audit.md
    ├── metadata-verification.md
    └── report-template.md
text
<installed-skill-dir>/
├── SKILL.md
├── scripts/
│   └── audit_latex_refs.py
└── references/
    ├── citation-claim-audit.md
    ├── metadata-verification.md
    └── report-template.md

Progressive Loading

渐进式加载

  • Always run
    scripts/audit_latex_refs.py
    for deterministic TeX/BibTeX/reference checks.
  • Read
    references/metadata-verification.md
    when checking DOI, arXiv, OpenReview, proceedings, or publisher metadata.
  • Read
    references/citation-claim-audit.md
    when the user asks whether citations support the claims made in the paper, or when doing a full pre-submission audit.
  • Use
    references/report-template.md
    for the final audit report.
  • 始终运行
    scripts/audit_latex_refs.py
    以执行确定性的TeX/BibTeX/引用检查。
  • 检查DOI、arXiv、OpenReview、会议集或出版商元数据时,阅读
    references/metadata-verification.md
  • 当用户询问引用是否支持论文中的声明,或进行全面的提交前审核时,阅读
    references/citation-claim-audit.md
  • 使用
    references/report-template.md
    生成最终审核报告。

Step 1 - Locate Paper Sources

步骤1 - 定位论文源文件

Determine:
  • paper root
  • main TeX file
  • all included TeX files
  • BibTeX files referenced by
    \bibliography{...}
    or
    \addbibresource{...}
  • target venue and submission mode if obvious
Useful local checks:
bash
find . -maxdepth 4 -name "*.tex" -o -name "*.bib"
find . -maxdepth 3 -name "main.tex" -o -name "paper.tex"
If the user provides a paper directory, use it. If no main file is provided, prefer
main.tex
, then
paper.tex
, then the TeX file containing
\begin{document}
.
确定:
  • 论文根目录
  • 主TeX文件
  • 所有被包含的TeX文件
  • 通过
    \bibliography{...}
    \addbibresource{...}
    引用的BibTeX文件
  • 若明确,目标会议/期刊及提交模式
有用的本地检查命令:
bash
find . -maxdepth 4 -name "*.tex" -o -name "*.bib"
find . -maxdepth 3 -name "main.tex" -o -name "paper.tex"
如果用户提供了论文目录,则使用该目录。若未提供主文件,优先选择
main.tex
,其次是
paper.tex
,最后是包含
\begin{document}
的TeX文件。

Step 2 - Run Deterministic Local Audit

步骤2 - 运行确定性本地审核

Run:
bash
python3 <citation-audit-skill-dir>/scripts/audit_latex_refs.py --paper-dir "$PAPER_DIR" --main "$MAIN_TEX"
Use an absolute path to the installed skill script. Do not assume a Claude-specific install path.
The script checks:
  • included TeX file discovery
  • citation keys in
    \cite
    ,
    \citet
    ,
    \citep
    ,
    \citealp
    ,
    \citeauthor
    ,
    \citeyear
    ,
    \textcite
    ,
    \parencite
  • BibTeX keys and basic syntax
  • missing citation keys
  • unused BibTeX entries
  • duplicate BibTeX keys
  • missing bibliography files
  • duplicate labels
  • undefined references
  • labels that are never referenced
  • unresolved LaTeX placeholders such as
    ??
    and citation placeholders such as
    [?]
If the script reports blocking issues, fix or report those before doing web metadata checks. Metadata validation is much less useful if the TeX/BibTeX graph is broken.
运行:
bash
python3 <citation-audit-skill-dir>/scripts/audit_latex_refs.py --paper-dir "$PAPER_DIR" --main "$MAIN_TEX"
使用已安装技能脚本的绝对路径。不要假设特定于Claude的安装路径。
该脚本检查:
  • 被包含TeX文件的识别情况
  • \cite
    \citet
    \citep
    \citealp
    \citeauthor
    \citeyear
    \textcite
    \parencite
    中的引用键
  • BibTeX键及基本语法
  • 缺失的引用键
  • 未使用的BibTeX条目
  • 重复的BibTeX键
  • 缺失的参考文献文件
  • 重复的标签
  • 未定义的引用
  • 从未被引用的标签
  • 未解析的LaTeX占位符(如
    ??
    )和引用占位符(如
    [?]
如果脚本报告阻塞性问题,在进行网络元数据检查之前先修复或报告这些问题。如果TeX/BibTeX关系图已损坏,元数据验证的用处会大打折扣。

Step 3 - Classify Findings

步骤3 - 分类检查结果

Use this severity model:
  • blocking
    : missing cited key, duplicate BibTeX key, undefined ref, duplicate label, invalid BibTeX structure, broken DOI for a cited work
  • important
    : metadata mismatch, wrong venue/year, likely duplicate entry, citation claim not supported, arXiv cited when peer-reviewed version should be cited
  • warning
    : unused BibTeX entry, unreferenced label, inconsistent key naming, missing optional DOI/URL
  • note
    : style cleanup, capitalization, field normalization, BibTeX key rename suggestion
Do not treat unused BibTeX entries as blocking unless the target venue or user requires a minimal bibliography.
使用以下严重程度模型:
  • blocking
    (阻塞性):缺失引用键、重复BibTeX键、未定义引用、重复标签、无效BibTeX结构、被引用作品的DOI损坏
  • important
    (重要):元数据不匹配、会议/年份错误、疑似重复条目、引用声明未被支持、应引用同行评审版本却引用了arXiv版本
  • warning
    (警告):未使用的BibTeX条目、未被引用的标签、键命名不一致、缺失可选DOI/URL
  • note
    (备注):格式清理、大小写规范、字段标准化、BibTeX键重命名建议
除非目标会议/期刊或用户要求最小化参考文献,否则不要将未使用的BibTeX条目视为阻塞性问题。

Step 4 - Verify Metadata

步骤4 - 验证元数据

Read
references/metadata-verification.md
.
For every cited key, verify the best available identifier:
  • DOI through publisher/CrossRef/doi.org
  • arXiv ID through arXiv
  • OpenReview URL or forum ID through OpenReview
  • proceedings URL for NeurIPS, ICML, ICLR, ACL Anthology, CVF, ACM, IEEE, Springer, or PMLR
Check:
  • title
  • author list or first author + author count
  • year
  • venue or publication status
  • DOI/arXiv/OpenReview/proceedings URL
  • whether a peer-reviewed version exists
When metadata cannot be verified, mark it explicitly instead of guessing.
阅读
references/metadata-verification.md
对每个引用键,验证最可靠的标识符:
  • 通过出版商/CrossRef/doi.org验证DOI
  • 通过arXiv验证arXiv ID
  • 通过OpenReview验证OpenReview URL或论坛ID
  • NeurIPS、ICML、ICLR、ACL Anthology、CVF、ACM、IEEE、Springer或PMLR的会议集URL
检查:
  • 标题
  • 作者列表或第一作者+作者数量
  • 年份
  • 会议/期刊或出版状态
  • DOI/arXiv/OpenReview/会议集URL
  • 是否存在同行评审版本
当元数据无法验证时,明确标记而非猜测。

Step 5 - Audit Citation Claims

步骤5 - 审核引用声明

Read
references/citation-claim-audit.md
for full guidance.
For each citation context, classify what the prose asks the citation to support:
  • background fact
  • prior method existence
  • closest related work
  • empirical result
  • theoretical result
  • dataset or benchmark
  • negative claim or limitation
  • comparison or state-of-the-art claim
Then check whether the cited paper actually supports that role. For high-risk claims, inspect the abstract, introduction, method/result section, and if needed the PDF.
High-risk contexts:
  • "first", "only", "state-of-the-art", "significantly", "provably", "guarantees"
  • claims about a paper's results or limitations
  • citations used to justify a baseline choice
  • citations used for theory assumptions
  • citations in contribution bullets or problem motivation
Do not silently rewrite scientific claims. If a citation does not support a claim, propose one of:
  • replace citation
  • weaken claim
  • add a more specific citation
  • move the claim to related work
  • mark as needing author confirmation
阅读
references/citation-claim-audit.md
获取完整指导。
对每个引用上下文,分类文本要求引用支持的内容类型:
  • 背景事实
  • 已有方法的存在性
  • 最相关的研究工作
  • 实证结果
  • 理论结果
  • 数据集或基准
  • 负面声明或局限性
  • 比较或最先进水平声明
然后检查被引用论文是否确实支持该角色。对于高风险声明,检查摘要、引言、方法/结果部分,必要时查看PDF。
高风险上下文:
  • "首次"、"唯一"、"最先进"、"显著地"、"可证明地"、"保证"
  • 关于论文结果或局限性的声明
  • 用于证明基线选择合理性的引用
  • 用于理论假设的引用
  • 贡献要点或问题动机中的引用
不要擅自改写科学声明。如果引用不支持声明,建议采取以下措施之一:
  • 替换引用
  • 弱化声明
  • 添加更具体的引用
  • 将声明移至相关工作部分
  • 标记为需作者确认

Step 6 - Fix Safe Issues

步骤6 - 修复安全问题

Safe auto-fixes:
  • add missing
    .bib
    extension resolution in the report
  • remove obvious duplicate BibTeX entries only after confirming they are truly identical
  • normalize capitalization braces in titles
  • add missing DOI/arXiv/URL fields when verified
  • fix BibTeX field spelling
  • rename labels or citation keys only if all TeX call sites are updated consistently
Never auto-fix:
  • citation claims whose support is ambiguous
  • substitution of one cited paper for another without explaining the scientific difference
  • venue status when multiple versions exist
  • author order if sources disagree
For any edit, keep the smallest possible diff.
可自动修复的安全问题:
  • 在报告中添加缺失的
    .bib
    扩展名解析
  • 仅在确认完全相同时删除明显重复的BibTeX条目
  • 标准化标题中的大小写括号
  • 验证后添加缺失的DOI/arXiv/URL字段
  • 修复BibTeX字段拼写错误
  • 仅在所有TeX调用位置都一致更新时,重命名标签或引用键
禁止自动修复:
  • 支持性不明确的引用声明
  • 未经解释科学差异就替换被引用论文
  • 存在多个版本时的会议/期刊状态
  • 来源不一致时的作者顺序
对于任何编辑,保持最小化的差异。

Step 7 - Write the Audit Report

步骤7 - 撰写审核报告

Use
references/report-template.md
.
The final report should include:
  • files checked
  • local TeX/BibTeX graph status
  • metadata verification status
  • citation-claim support status
  • blocking fixes required before submission
  • recommended non-blocking cleanup
  • unresolved items requiring author judgment
If the user asks for a saved report and gives no path, use:
text
docs/reports/citation_audit_YYYY-MM-DD.md
使用
references/report-template.md
最终报告应包含:
  • 已检查的文件
  • 本地TeX/BibTeX关系图状态
  • 元数据验证状态
  • 引用声明支持状态
  • 提交前需修复的阻塞性问题
  • 建议的非阻塞性清理工作
  • 需作者判断的未解决事项
如果用户要求保存报告但未提供路径,使用:
text
docs/reports/citation_audit_YYYY-MM-DD.md

Step 8 - Final Sanity Check

步骤8 - 最终 Sanity 检查

Before finalizing:
  • all cited keys resolve to exactly one BibTeX entry
  • all required TeX references resolve to exactly one label
  • blocking citation/reference problems are tracked as actions when project memory exists
完成前检查:
  • 所有引用键都能解析到恰好一个BibTeX条目
  • 所有必需的TeX引用都能解析到恰好一个标签
  • 当存在项目记忆时,阻塞性引用/引用问题被列为待办事项

Step 9 - Write Back to Project Memory

步骤9 - 写入项目记忆

If the project uses
research-project-memory
, update:
  • memory/risk-board.md
    : blocking or important citation, metadata, label, reference, or citation-claim risks
  • memory/action-board.md
    : concrete fixes for missing keys, metadata corrections, unsupported claims, or broken refs
  • memory/claim-board.md
    : claims that must be weakened because citations do not support them
  • paper/.agent/paper-status.md
    : citation-audit status and unresolved author decisions
Use
observed
for deterministic TeX/BibTeX graph findings and
needs-verification
for metadata or claim-support issues not fully checked.
  • every blocking metadata issue is fixed or explicitly listed
  • high-risk citation claims have been audited
  • unresolved citation correctness questions are not hidden
  • the final answer distinguishes deterministic script findings from web/semantic verification findings
如果项目使用
research-project-memory
,更新:
  • memory/risk-board.md
    :阻塞性或重要的引用、元数据、标签、引用或引用声明风险
  • memory/action-board.md
    :针对缺失键、元数据修正、未被支持的声明或损坏引用的具体修复措施
  • memory/claim-board.md
    :因引用不支持而必须弱化的声明
  • paper/.agent/paper-status.md
    :引用审核状态及未解决的作者决策
对确定性的TeX/BibTeX关系图检查结果使用
observed
(已观测),对未完全检查的元数据或声明支持问题使用
needs-verification
(需验证)。
  • 所有阻塞性元数据问题已修复或明确列出
  • 高风险引用声明已审核
  • 未解决的引用正确性问题未被隐藏
  • 最终答案区分确定性脚本检查结果与网络/语义验证结果