AI-powered bibliographic reference verification skill
Verify the authenticity, accuracy, and formatting of your references — powered by CrossRef API and web search.
English | 中文
ref-check is a Claude Code skill that verifies bibliographic references against authoritative sources and generates a professional Word (.docx) verification report. It catches real problems that manual checking would miss — fabricated citations, outdated preprints, broken DOIs, author attribution errors, and formatting inconsistencies.
- Dual-layer verification: CrossRef API for structured metadata lookup, then web search for secondary confirmation. No reference is judged on a single source alone.
- Multi-type support: Journal articles, preprints (arXiv/SSRN/bioRxiv), books, conference papers, reports, blog posts, webpages, datasets, theses, and more.
- AI fabrication detection: Identifies suspected AI-generated fake references by cross-referencing multiple signals — not just "can't find it."
- Citation style checking: APA 7th, GB/T 7714, Vancouver, Nature, IEEE, and more. Detects specific deviations (missing DOI, wrong author format, punctuation errors).
- Chinese reference support: Searches CNKI, Wanfang, and VIP databases for Chinese-language (中文) references that aren't indexed in CrossRef.
- Structured Word report: Executive summary with status counts, per-reference details with color-coded status markers (✅
⚠️ ❌ℹ️), suggested corrections, and optional appendices. - Socratic workflow: Asks about your target journal, preferred citation style, and report language before starting — so you get exactly the report you need.
┌─────────────────┐
│ Extract refs │ Parse references from Word, PDF, or text
└────────┬────────┘
│
┌────────▼────────┐
│ Ask the user │ Report language? Target journal? Citation style?
└────────┬────────┘
│
┌────────▼────────┐
│ CrossRef API │ Primary verification via api.crossref.org
└────────┬────────┘
│
┌────▼────┐
│ Match? │
└────┬────┘
Yes │ No / Partial
│ │
│ ┌─▼──────────────┐
│ │ Web search │ Secondary verification via online search
│ └─┬──────────────┘
│ │
┌───▼────▼────────┐
│ Final verdict │ ✅ Verified / ⚠️ Attention / ❌ Problem / ℹ️ Unverifiable
└────────┬────────┘
│
┌────────▼────────┐
│ Generate report │ Word (.docx) with consistent formatting
└─────────────────┘
- Python 3.8+ with
python-docxinstalled:pip install python-docx
Global install (available in all projects):
git clone https://github.com/CingyQ/ref-check.git
cp -r ref-check/ref-check ~/.claude/skills/ref-checkProject-level install (only for the current project):
git clone https://github.com/CingyQ/ref-check.git
cp -r ref-check/ref-check .claude/skills/ref-checkAfter copying, restart Claude Code or start a new session for the skill to be detected.
Clone the repository and reference the skill in your AGENTS.md:
git clone https://github.com/CingyQ/ref-check.git ~/.codex/skills/ref-checkThen add to your AGENTS.md:
## Skills
- ref-check: ~/.codex/skills/ref-check/ref-check/SKILL.md — Verify bibliographic referencesClone the repository to your preferred skills directory:
git clone https://github.com/CingyQ/ref-check.git /path/to/skills/ref-checkPoint your assistant's skill configuration to ref-check/ref-check/SKILL.md.
- Download or clone this repository
- Copy the
ref-check/directory into your Claude Code skills folder:- Global:
~/.claude/skills/ref-check/ - Project:
.claude/skills/ref-check/
- Global:
Simply ask Claude to check your references:
Check the references in my paper.docx
Please verify my bibliography — I'm submitting to Nature.
核查一下这篇文档里的参考文献
Claude will:
- Extract all references from your document
- Ask you about report language, target journal, and citation style
- Verify each reference via CrossRef API + web search
- Generate a Word report with findings
Input: A manuscript with 5 references ready for submission to an APA-formatted journal.
1. Vaswani, A., Shazeer, N., Parmar, N., et al. Attention is all you need.
Advances in Neural Information Processing Systems 30, 5998–6008 (2017).
2. Brown, T. B., Mann, B., Ryder, N., et al. Language models are few-shot learners.
arXiv preprint arXiv:2005.14165 (2020).
3. Wei, J., Wang, X., Schuurmans, D., et al. Chain-of-thought prompting elicits
reasoning in large language models. arXiv preprint arXiv:2201.11903 (2022).
4. Zhang, L. & Chen, R. Multi-modal fusion networks for sentiment analysis.
Journal of Computational Linguistics 45, 112–130 (2023).
5. World Health Organization. Global health estimates 2024.
https://www.who.int/data/global-health-estimates (2024).
What ref-check found:
| # | Reference | Status | Finding |
|---|---|---|---|
| 1 | Vaswani et al., NeurIPS 2017 | ✅ Verified | CrossRef confirmed. All metadata matches. |
| 2 | Brown et al., arXiv 2020 | Paper was published at NeurIPS 2020 — should cite the conference version instead of the preprint. | |
| 3 | Wei et al., arXiv 2022 | Paper was published at NeurIPS 2022 — should cite the conference version. | |
| 4 | Zhang & Chen, 2023 | ❌ Problem | No record found in CrossRef or web search. "Journal of Computational Linguistics" does not exist — the real journal is "Computational Linguistics" (MIT Press). Suspected AI-fabricated reference. |
| 5 | WHO, 2024 | ✅ Verified | URL is accessible. Content matches the citation. |
Additional findings:
- References 1–3 are in Nature/Vancouver style, not APA 7th. Missing DOIs, incorrect author format, and year placement need correction.
- Suggested APA 7th corrections are provided for each reference in the report.
Output: A Word (.docx) report with executive summary, per-reference verification details, suggested corrections, and formatting analysis appendix.
ref-check/
├── ref-check/
│ ├── SKILL.md # Main skill definition
│ ├── scripts/
│ │ └── crossref_lookup.py # CrossRef API query tool
│ └── references/
│ └── report-spec.md # Word report formatting specification
├── README.md # English documentation
├── README_zh.md # Chinese documentation
└── LICENSE
The bundled crossref_lookup.py script queries CrossRef — the largest DOI registration agency, covering 150M+ scholarly records. It supports:
- Title + author search: Finds the best match and returns a confidence score
- DOI validation: Confirms a DOI resolves and returns correct metadata
- Batch mode: Processes multiple references efficiently with polite rate limiting
When CrossRef is insufficient (non-English works, blogs, reports, webpages), the skill falls back to web search to:
- Confirm the work exists on its official platform
- Check URL accessibility for web-based references
- Search Chinese databases (CNKI, Wanfang, VIP) for Chinese-language references
The skill is deliberately conservative:
- "Not found" ≠ "fabricated" — many real works aren't in CrossRef
- AI fabrication is only flagged when multiple red flags co-occur
- Unverifiable references (internal reports, personal communications) are labeled honestly, not forced into pass/fail
Issues and pull requests are welcome. If you find a reference type that isn't handled well, please open an issue with an example.