review-23-reference-curator
Curate, verify, insert, format, and audit high-quality references for medical SCI review manuscripts. Use when the user asks to find references, add citations to Chinese or English academic text, repl
SKILL.mdskills/review-23-reference-curator/SKILL.md
GitHubReview 23 Reference Curator
Overview
Use this skill before review-17-citation-audit when references are the central task. The goal is to convert manuscript claims into a traceable, high-quality reference set that is recent, verifiable, journal-appropriate, and directly matched to the text.
This skill may borrow process standards from Nature/Science/CNS citation skills, academic search skills, citation verifier skills, and the local review workbench, but medical evidence authority always comes from verified literature and databases, not from style skills.
Position In The Review Workbench
- Use
review-03-search-stringandreview-04-search-executionfor broad systematic searches. - Use this skill for paragraph-level citation insertion, reference replacement, DOI/PMID completion, and reference-quality screening.
- Use
review-11-evidence-matrixwhen extracted evidence needs to support manuscript arguments across sections. - Use
review-17-citation-auditafter this skill to confirm final numbering, DOI/PMID integrity, and claim-citation alignment. - Use
review-22-skill-routerwhen combining this workflow with Nature, Science, academic, PRISMA, or other external skill systems.
Non-Negotiable Rules
- Do not invent references, DOI, PMID, author names, title, journal, year, volume, issue, pages, article number, or findings.
- Prefer 2021-2026 references. Include 2020 only when the user allows it or when it is a high-impact guideline, consensus, landmark trial, major review, or foundational method paper.
- Require both DOI and PMID for core biomedical references. If either is missing, mark it as
exception candidateand explain why it may or may not be acceptable. - Verify each reference through PubMed plus DOI/Crossref or publisher page when possible. Use Web of Science if available. Use Google Scholar only as a discovery aid or secondary cross-check.
- Prefer JCR Q1-Q2, CAS high-partition journals, field-defining journals, guidelines, consensus statements, high-quality systematic reviews/meta-analyses, and major original studies.
- Avoid MDPI, Frontiers, Scientific Reports, Medicine, and other user-excluded or quality-risk sources unless no credible alternative exists and the user explicitly approves the risk.
- Match each citation to the exact claim. Do not cite a paper merely because its topic is similar.
- Do not use correlation evidence to support causal language.
- Do not use basic or mechanistic evidence as direct support for clinical efficacy unless the text clearly frames it as mechanistic or hypothesis-generating.
- Do not use review articles to support precise numeric results when a primary study, guideline, or dataset is needed.
Source Priority
Use this order for biomedical review manuscripts:
- PubMed record and PMID.
- Publisher page and DOI resolver/Crossref metadata.
- Guidelines, consensus, Cochrane, major society statements, and clinical trial registrations where appropriate.
- Web of Science, Scopus, Semantic Scholar, or OpenAlex for citation context and metadata cross-checking.
- Google Scholar only for discovery, citation chasing, or locating otherwise missing official pages.
When using external skill layers:
nature-citation: use for strict segmentation, support grading, and CNS/Nature-family candidate discovery; do not let it override the user's excluded-source list.nature-polishing: use only after evidence is settled, for prose clarity and citation discipline.academic-mcp-search: use for structured metadata, deduplication, DOI-first merging, and audit-ready tables.citation-verifier: use for local placeholder scans, duplicate keys, DOI/PMID gaps, and final bibliography hygiene.
Workflow
1. Segment The Text
- Preserve the user's original paragraph order.
- Treat every substantive paragraph as needing at least one citation unless it is purely transitional.
- Split long paragraphs into focused citable claims when they contain multiple mechanisms, clinical statements, epidemiology facts, or methodological assertions.
- Assign stable IDs:
P001,P002for paragraphs andC001,C002for individual claims.
2. Classify Each Claim
For each claim, assign one type:
backgroundepidemiologymechanismclinical diagnosisclinical treatmentprognosisbiomarkerbioinformatics methodguideline or consensusreview context
Then identify population/model, disease, exposure/intervention, molecular pathway, endpoint, and whether the wording implies causality.
3. Search And Screen
- Translate Chinese claims into English search concepts while preserving the original scientific meaning.
- Search with precise terms first, then synonyms and broader terms.
- Prioritize recent high-quality evidence from 2021-2026.
- Keep a dropped-records log for excluded papers, especially if excluded for year, no PMID, no DOI, low-quality journal, wrong claim support, or user-excluded source.
- For clinical claims, prioritize guidelines, consensus statements, RCTs, large cohorts, systematic reviews, and meta-analyses over small retrospective studies.
- For mechanism claims, prioritize original mechanistic studies in appropriate models, supported by high-quality reviews only for context.
- For bioinformatics claims, prioritize method papers, dataset papers, benchmark studies, and reproducible workflows with stable identifiers.
4. Verify Each Candidate
For every retained candidate, check:
- Title matches DOI/PubMed metadata.
- DOI resolves or appears in Crossref/publisher metadata.
- PMID exists and matches title, journal, year, and authors.
- Journal is acceptable for the target review.
- Publication year fits the user's time window.
- Article type fits the claim.
- Abstract or full text supports the exact claim.
- No obvious retraction, expression of concern, or major mismatch is visible.
Support grades:
strong: directly supports the exact claim.partial: supports only part of the claim or a narrower setting.background: supports field context but not the exact assertion.weak: thematically related but not adequate as direct support.reject: unverifiable, mismatched, outdated without justification, excluded source, or poor claim fit.
Only use strong, partial, or clearly labeled background citations in the manuscript.
5. Insert Citations
- Insert numeric citations in the target journal style, usually
[1],[2]. - Keep numbering in first-appearance order.
- Use the smallest adequate number of citations per claim.
- Place citations close to the supported claim, usually at the end of the sentence or clause.
- If one paragraph contains distinct factual claims, place separate citations near each claim instead of stacking all references at the paragraph end.
- Do not change the scientific meaning while inserting citations.
6. Format References
- Default to the target journal's numbered style. For World Journal of Gastroenterology, use numbered references and verify the latest official author/title/journal/year/volume/page/PMID/DOI requirements at use time.
- If the user requests Nature or Science style, follow the requested journal family format after verifying current instructions.
- Preserve DOI and PMID in the output whenever allowed by the target format or in the audit table when not shown in the reference list.
- Do not fabricate missing volume, issue, page, article number, DOI, or PMID.
7. Produce Audit Artifacts
Always provide a compact audit table with:
- Reference number.
- Supported paragraph/claim ID.
- Claim summary.
- First author and year.
- Journal.
- Article type.
- PMID.
- DOI.
- Support grade.
- Source-quality note.
- Risk or exception.
For larger manuscripts, also recommend saving:
reference_candidates.csvclaim_citation_check.csvcitation_crossref_pubmed_check.csvcitation_audit_report.md
Output Format
Unless the user requests otherwise, return:
For Chinese users, write the visible response in Chinese while keeping table field names stable when useful.
Task decision: reference insertion, replacement, verification, formatting, or final audit preparation.Claim analysis: paragraph or claim IDs with claim type and citation need.Revised text with citations: manuscript text with numeric citations.Reference list: target journal style.Citation audit table: DOI/PMID/year/journal/support-grade/risk table.Risks and gaps: missing DOI/PMID, no direct evidence, old but necessary references, or excluded-source conflicts.
Stop Conditions
- If no reference satisfies DOI, PMID, year, journal quality, and claim support requirements, state that clearly and propose a narrower search strategy.
- If a requested claim appears unsupported or overclaimed, suggest safer wording rather than forcing a citation.
- If the user asks to cite a low-quality or excluded source, flag the source risk and provide higher-quality alternatives first.
Reference File
Read references/reference-quality-gates.md when the task requires strict inclusion/exclusion decisions, a reference audit table, or target-journal-ready citation insertion.