Nature Citation

Nature Citation — Router

This skill is split into two layers:

• A static layer under static/ that holds versioned, reusable content fragments (core principles and scope, the Chinese-user operating mode, and the citation workflow).
• A dynamic layer (this file plus manifest.yaml) that loads the core every time and reaches for heavier material only when a step needs it.

Do not try to apply the citation logic from memory or from this router. Always load fragments from disk as described below.

Routing protocol

Follow these four steps every time the skill is invoked.

1. Load the manifest and the core layer

Read manifest.yaml. Then read every file listed under always_load:

• static/core/principles.md — what the skill produces, the strict journal scope, the source hierarchy, and the search-quality rules.
• static/core/chinese-mode.md — how to operate when the user writes in Chinese or asks for Nature系列/CNS及子刊 style support.
• static/core/workflow.md — the seven-step workflow and the final report format.

2. No content axis — confirm scope and language inline

Unlike the other nature-* skills, nature-citation has no fragment axis. Its variation is runtime parameters, not different content bodies:

• journal scope — Nature系列 / CNS / CNS及子刊 / flagship-only. Read it from the user's wording (see core/principles.md) and pass it to the script as --scope.
• user language — if the user writes Chinese, follow core/chinese-mode.md (Chinese notes, English search queries).
• input length — if there are more than ~10 segments, switch to the batched long-article strategy in references/script-usage.md.

State the detected scope and date limits in one short line before searching.

3. Run the workflow

Follow the seven steps in core/workflow.md: segment, parse, search, evaluate support conservatively, export one reference-manager file, generate review artifacts when useful, and report with the HTML browser path first. Prefer scripts/nature_citation.py for the search/export when internet access is available; open references/script-usage.md for its full flag list and the long-article batch strategy.

Never present a paper as support merely because its title is related, and never cite a metadata-only candidate without checking the abstract or publisher page. Do not invent missing bibliographic fields.

4. Reach for references only when needed

The files under references/ are deep references, not defaults. Open them on demand per the references.on_demand table in the manifest:

• running the script, full flags, long-article batching → references/script-usage.md.
• turning a claim into search queries and support grades → references/search-strategy.md.
• the exact Nature/CNS journal-family boundary → references/journal-scope.md.
• RIS / EndNote / Zotero RDF export details → references/ris-endnote.md.

Why this split

• The static layer is versioned and reviewable; the core stays small for a normal short run.
• The dynamic layer keeps each invocation cheap: the script flag dump and long-article strategy load only when actually running a search.
• The router itself is short on purpose. Update fragments and references, not this file, when adding scope.
• This structure mirrors nature-writing, nature-polishing, nature-reader, nature-paper2ppt, and nature-figure.

nature-citation skill

A citation-search skill for turning manuscript text or standalone claims into strict Nature / CNS-family reference exports with segment-level mapping and reference-manager-ready downloads.

This skill is bilingual-aware. It accepts Chinese manuscript text and citation requests such as "分段引用", "Nature系列引用", "CNS及子刊", "补引用", "支撑文献", or "导出 Zotero", then searches with English scientific concepts while returning Chinese review notes by default.

What it does

• splits manuscript text into citable segments with stable IDs such as S001, S002, and S003
• converts each segment into search queries for Crossref-led discovery
• filters results to Nature Portfolio, the AAAS Science family, Cell Press, or flagship-only scope
• maps each segment to candidate citations and suggested in-text insertion markers
• exports one reference-manager file in ENW, RIS, or Zotero RDF
• optionally builds JSON, TSV, Markdown, and HTML review artifacts for manual screening
• supports long-article batch processing with partial checkpoints
• retries transient Crossref failures instead of failing immediately
• supports limiting one run to part of a long manuscript
• supports DOI-only export when the user already knows which records should be included

Source hierarchy

• Crossref structured metadata and DOI records
• PubMed / NCBI E-utilities for biomedical cross-checking when relevant
• Official publisher pages from Nature Portfolio, AAAS Science, and Cell Press
• Secondary scholarly indexes only as discovery aids, never as the sole support basis

File structure

The skill uses a router/static-dynamic split (like the other nature-* skills): a short SKILL.md router plus a manifest.yaml. nature-citation is a linear workflow with no content axis, so the split is core (always loaded) plus on-demand references.

nature-citation/
├── SKILL.md                     # short router
├── manifest.yaml                # always_load core + on-demand references (no axis)
├── README.md
├── static/
│   └── core/                    # always loaded
│       ├── principles.md        # what it produces, journal scope, source hierarchy, search rules
│       ├── chinese-mode.md      # Chinese-user operating mode
│       └── workflow.md          # the 7-step workflow + report format
├── references/                  # opened on demand
│   ├── script-usage.md          # nature_citation.py flags + long-article batch strategy
│   ├── journal-scope.md
│   ├── ris-endnote.md
│   └── search-strategy.md
└── scripts/
    └── nature_citation.py

When to use

• adding citations to a paragraph, abstract, introduction, results, or discussion section
• turning long text into segment-by-segment citation candidates
• restricting references to Nature系列, CNS, CNS及其子刊, or 只看正刊
• exporting references for EndNote, Zotero, or other citation managers
• screening whether a sentence has direct support, partial support, or only background support
• producing an HTML review page where the user filters by year, selects citations, and downloads only the records they want

Long-text behavior

This skill now has a safer path for long inputs such as a full Introduction or multi-paragraph text.

• for short inputs, it still works as a normal one-pass citation search
• for longer inputs, it can process segments in batches
• after each batch, it writes a partial export checkpoint so progress is not lost if a later batch fails
• transient Crossref failures are retried automatically

Useful rules of thumb:

• 1-10 segments: normal run
• 11-25 segments: prefer batch mode
• 26+ segments: prefer section-by-section runs

Design intent

The skill should prioritize defensibility over volume. It is designed to help the user find likely in-scope papers, not to pretend that metadata alone proves a claim. Every exported record should preserve real metadata, avoid fabricated fields, and make the evidence-review burden explicit.

For long manuscripts, the design goal is not only citation quality but also run stability: fewer lost runs, smaller batches, and a reviewable checkpoint trail.

Reference map

• search-strategy.md: claim decomposition, support grades, and common retrieval failure modes
• journal-scope.md: Nature / Science / Cell family boundaries and flagship-only interpretation
• ris-endnote.md: ENW, RIS, and Zotero RDF export guidance
• scripts/nature_citation.py: local CLI for segmentation, Crossref retrieval, export, and HTML review generation

Useful CLI options

• --batch-size 2: process long text in smaller batches
• --max-segments 12: cap the number of segments processed in one run
• --max-retries 2: retry transient Crossref failures
• --sleep 0.3: shorter default pause between requests
• --with-artifacts: generate HTML, TSV, JSON, and Markdown review files

Notes

• Default output is a single reference-manager file; additional artifacts are opt-in.
• metadata-only candidate means the abstract or full text still needs human review before citation.
• The HTML review page can export selected references as ENW, RIS, or Zotero RDF.
• For long texts, --with-artifacts is strongly recommended because the HTML browser is the easiest way to curate results.
• Batch mode writes .partial.enw / .partial.ris / .partial.rdf checkpoints during the run before the final export is written.

Journal Scope

The skill's default journal-family boundary is intentionally practical rather than exhaustive. Use it to find likely Nature/CNS-family candidates, then verify exact journal status on official pages if the author needs a strict portfolio definition.

Default families

Nature Portfolio

Include:

• Nature
• journals beginning with Nature , such as Nature Medicine, Nature Biotechnology,

Nature Methods, Nature Materials, Nature Genetics, Nature Communications

• Communications journals, such as Communications Biology, Communications Chemistry,

Communications Materials, Communications Earth & Environment, Communications Medicine

• npj journals
• Scientific Reports

Be careful with unrelated titles that include the common word "nature".

Science family

Include by default:

• Science
• Science Advances
• Science Translational Medicine
• Science Signaling
• Science Immunology
• Science Robotics

The AAAS Science Partner Journal program is not included by default unless the user asks for partner journals or broader AAAS coverage.

Cell Press

Include the flagship Cell, major primary-research Cell Press journals, Cell Reports titles, and Trends review journals. The local script recognizes common Cell Press titles and any title beginning with Trends in .

Because Cell Press launches and reorganizes titles over time, verify official pages for exhaustive coverage or a current journal list.

Flagship-only scope

Use only:

• Nature
• Science
• Cell

This is appropriate when the user says "只看正刊", "主刊", "flagship only", or explicitly excludes subjournals.

Official source notes

• Crossref REST API can retrieve scholarly metadata, search works, and filter exact fields such as

container-title and issn.

• NCBI E-utilities provide structured access to PubMed and other Entrez databases; observe request

frequency guidance.

• EndNote documents Reference Manager (RIS) as an import option for RIS files.
• Nature Portfolio, AAAS, and Cell Press official pages should be checked when exact current journal

coverage matters.

RIS, EndNote, and Zotero RDF Output

EndNote can import RIS files using the Reference Manager (RIS) import option. Use .ris as the default exchange format because it is plain text, widely supported, and easy to inspect.

RIS mapping for journal articles

Use these tags:

TY  - JOUR
TI  - Article title
AU  - Author, Given
T2  - Journal title
JO  - Journal title
PY  - Publication year
Y1  - YYYY/MM/DD when available
VL  - Volume
IS  - Issue
SP  - First page or article number
EP  - Last page
DO  - DOI
UR  - URL
SN  - ISSN
N2  - Abstract or short metadata note, only when safely available
ER  -

Rules:

• Write one AU line per author.
• Use TY - JOUR for journal articles.
• End every record with ER -.
• Do not invent missing fields.
• Prefer DOI over URL when both exist.
• Keep notes concise; avoid copying long abstracts into RIS unless the source terms allow it.

EndNote import instruction

Tell the user:

In EndNote: File > Import > File, choose the `.ris` file, set Import Option to
Reference Manager (RIS), then import.

Menu labels vary slightly by EndNote version and operating system, so avoid over-specific UI claims unless the user gives their exact EndNote version.

Zotero RDF guidance

Use .rdf when the user explicitly asks for Zotero import/export.

Preferred structure:

<rdf:RDF ...>
  <bib:Article rdf:about="https://doi.org/...">
    <z:itemType>journalArticle</z:itemType>
    <dcterms:isPartOf rdf:resource="urn:..."/>
    <bib:authors>...</bib:authors>
    <dc:title>...</dc:title>
    <dc:date>YYYY-MM-DD</dc:date>
    <dc:identifier>...</dc:identifier>
    <bib:pages>...</bib:pages>
    <z:citationKey>...</z:citationKey>
  </bib:Article>
  <bib:Journal rdf:about="urn:...">...</bib:Journal>
</rdf:RDF>

Rules:

• Export one bib:Article per citation.
• Represent authors as foaf:Person nodes inside rdf:Seq.
• Deduplicate journal container nodes by journal/ISSN/volume/issue identity.
• Do not invent abstracts, attachments, or fields that are not present in metadata.

Script usage and long-article strategy

Open this reference when running scripts/nature_citation.py or when the input is long enough to need batching. It expands step 3 of the workflow.

Running the script

Start with scripts/nature_citation.py when internet access is available:

python scripts/nature_citation.py \
  --text "PASTE MANUSCRIPT TEXT HERE" \
  --scope cns \
  --outdir /tmp/nature-citation \
  --format enw \
  --with-artifacts

Useful options

• --text-file manuscript.txt: read long text from a file.
• --claim "CLAIM TEXT" or --claim-file claims.txt: treat each claim as a segment.
• --doi 10.xxxx/xxxxx or --doi-file dois.txt: export known DOI records after screening.
• --scope nature: Nature Portfolio-style journals only.
• --scope flagship: Nature, Science, and Cell only.
• --from-year 2018 --to-year 2026: constrain publication dates.
• --rows 40: raise for broad searches; keep top candidates manageable.
• --per-segment 3: number of citation candidates to keep per segment.
• --max-retries 2: retry transient Crossref failures before skipping a query.
• --format enw|ris|zotero-rdf: export format. If omitted and --output-file is set, infer from suffix.
• --mailto you@example.com: use Crossref's polite pool.
• --batch-size 10: process segments in batches of N. Each batch writes an incremental export file.
• --max-segments 20: only process the first N segments. Useful for testing or section-by-section workflows.
• --sleep 0.3: seconds between Crossref requests. Default is 0.3; raise to 1.0 if rate-limited.

Long-article strategy

When the input text is longer than roughly 3000 characters (about 10+ segments), switch to a batched workflow to avoid timeout, context overflow, or incomplete results:

1. Auto-detect length. Count segments after segmentation. If there are more than 10 segments, switch to batch mode automatically. 2. Split by section. Prefer splitting at paragraph double-line breaks or explicit section headings (Introduction, Results, etc.) so each batch is a coherent unit, not arbitrary sentence groups. 3. Process each batch independently. Run the script once per batch using --batch-size or --max-segments, OR split the text externally and call the script once per chunk. Each call writes its own intermediate export file. 4. Merge results at the end. After all batches finish, combine the intermediate files into one final export. Deduplicate by DOI. 5. Minimize inline analysis. For long articles, do NOT write detailed support-grade notes for every single segment inline. Instead:

• Write a compact summary table (segment ID → best candidate → support grade).
• Point the user to the HTML visualization for full browsing.
• Only elaborate on segments where no candidate was found or evidence is contradictory.

Quick guide

For long texts, prefer the HTML browser for review and selection instead of relying only on inline notes.

Search Strategy

Turn claims into searchable concepts

Break each sentence into:

• phenomenon: what is being claimed
• entity: gene, protein, pathway, compound, intervention, technology, population, or ecosystem
• relationship: increases, decreases, predicts, regulates, causes, associates with, improves, detects
• context: species, tissue, disease, cell type, geography, time period, device, method, or dataset
• boundary: "in cancer cells", "after treatment", "in older adults", "under drought", etc.

Create search queries at three levels:

1. precise: entity + relationship + outcome + context 2. synonym: alternate names and abbreviations 3. broad: field context if no direct paper is found

For Chinese claims, translate the scientific concepts, not the sentence literally. Keep acronyms and standard nomenclature unchanged.

Support grading

Use the smallest support grade that is defensible:

Evidence note template

Claim: [original claim]
Paper: [first author/year/title/journal/DOI]
Support grade: [grade]
Evidence basis: [title/abstract/publisher page/full text]
Reasoning: [why the result supports or does not support the exact claim]
Citation wording: [how to phrase the manuscript sentence if using this citation]

Common failure modes

• The paper is related to the same disease but tests a different mechanism.
• The paper supports an association, but the manuscript sentence claims causality.
• The evidence is in a different species, cell type, or clinical population.
• A review is used as primary evidence when original research exists.
• The claim is too broad for a single citation.
• The searched journal title contains "Nature" but is not a Nature Portfolio journal.

Better search moves

• Add the method or model when results are broad: single-cell, CRISPR screen, organoid,

randomized, cohort, meta-analysis, cryogenic electron microscopy.

• Add context terms when there are many irrelevant hits: tissue, species, cell type, disease subtype,

exposure, intervention, or outcome.

• Search the opposite direction if the claim might be overconfident: inhibits vs activates,

resistance vs sensitivity, risk vs protective.

• Use recent limits for fast-moving areas, but remove them if no direct CNS/Nature-series paper appears.

Chinese-user operating mode

When the user writes in Chinese, asks for "Nature系列", "CNS及其子刊", "支撑文献", "补引用", "自动给出引用", "分段引用", "导出EndNote", "RIS", "Zotero", "RDF", or provides Chinese manuscript text:

• Accept the text in Chinese, but search using English concept queries unless the topic is explicitly China-specific or Chinese-language scholarship.
• Return segment notes and evidence notes in Chinese by default.
• Preserve the exact source segment and translate it into one or more English search claims.
• Flag overclaiming clearly in Chinese: 强支撑, 部分支撑, 背景支撑, 不建议引用为该句支撑.
• Do not present a paper as supporting the claim merely because its title is related.

Core principles (citation)

Use this skill to turn manuscript text into a defensible citation export:

• segmented text with citation candidates for each segment
• a reference-manager import file in .enw, .ris, or Zotero .rdf
• conservative evidence notes explaining whether each candidate truly supports the segment

Default scope

Interpret journal scope from the user's wording, but keep the filter strict:

• Nature系列: search Nature Portfolio first. Include Nature, Nature [field], Nature Communications, Communications [field], Scientific Reports, and npj journals.
• CNS: search Cell, Nature, and Science plus their major sister journals.
• CNS及其子刊 or CNS/sister journals: search only accepted flagship and subjournal titles in Nature Portfolio, the AAAS Science family, and Cell Press.
• 只要Nature/Science/Cell正刊: restrict to the flagship journals Nature, Science, and Cell.

Do not treat merely related journals as in-scope. A title is valid only if it is in the accepted publisher-family whitelist or clearly matches the official naming pattern for that family. If the user needs an exhaustive or submission-critical boundary, verify current official journal pages before finalizing because journal portfolios change. The exact boundary and official source notes are in references/journal-scope.md.

Source hierarchy

Use sources in this order:

1. Structured bibliographic metadata: Crossref, PubMed/NCBI E-utilities, DOI metadata. 2. Publisher pages: nature.com, science.org, cell.com, and official journal pages. 3. Full text or abstract pages, if accessible. 4. Secondary databases such as Google Scholar, Semantic Scholar, Web of Science, or Scopus only as discovery aids, not as the sole support basis.

Prefer structured APIs for metadata and publisher pages for claim verification. If metadata and publisher page disagree, preserve the DOI and journal-page facts and flag the discrepancy.

Search quality rules

• Prefer precision over volume. A useful answer is usually 3-8 candidates, not 50 loosely related papers.
• Use exact phrase searches only for distinctive terms; otherwise use concept terms and synonyms.
• Check journal identity. Many journals contain the word "nature" but are not Nature Portfolio journals.
• Treat citation count as a tie-breaker, not evidence of support.
• Capture retractions, corrections, and expressions of concern when visible in Crossref or publisher metadata.
• Date-sensitive topics require current searching and an explicit search date.
• For medical, clinical, or safety claims, search current literature and state that citations do not replace clinical guidance or systematic review.

Source notes

This skill is based on public bibliographic APIs and official publisher/import documentation: Crossref REST API and filters, NCBI E-utilities, EndNote RIS import options, Nature Portfolio, AAAS Science journals, and Cell Press portfolio descriptions. Verify pages at use time when exact journal coverage or current import behavior matters.

Workflow

Run these seven steps for any citation job. For runs with more than ~10 segments, switch to the batched long-article strategy in references/script-usage.md.

1. Segment the text

• Split long text into citable segments. Prefer paragraph boundaries first, then sentence boundaries.
• Keep each segment focused on one citable idea when possible.
• Preserve original order and stable segment IDs such as S001, S002, S003.
• Skip obvious non-citable connective sentences unless the user asks to cite every sentence.
• For very long text, process in batches but keep a single final mapping table.

Default segmentation rules: use blank lines as paragraph boundaries; if a paragraph is longer than about 700 characters or contains multiple claims, split into sentences; merge very short fragments into neighboring text unless they contain a distinct claim; keep section headings as labels, not as citable segments. If the input has more than about 10 segments, prefer batch mode.

2. Parse each segment

For each citable segment:

• Extract the core claim in one sentence.
• Identify claim type: mechanism, association, method, clinical, epidemiology, background, definition, or review-context.
• Identify entities, intervention/exposure, outcome, population/model, directionality, and boundary.
• Convert the claim into 2-4 English search queries: one precise query with all key terms; one synonym query; one broader background query; one methods or model query if relevant.

If the claim is too broad, split it into citable subclaims rather than searching the whole sentence. For deeper help turning a claim into queries and support grades, open references/search-strategy.md.

3. Search candidate papers

Prefer scripts/nature_citation.py when internet access is available. The full flag list, polite-pool/rate-limit options, and the long-article batch strategy are in references/script-usage.md. A minimal run:

python scripts/nature_citation.py \
  --text "PASTE MANUSCRIPT TEXT HERE" \
  --scope cns \
  --outdir /tmp/nature-citation \
  --format enw \
  --with-artifacts

When the topic is biomedical or PubMed-indexed, also search PubMed with journal filters and compare results against Crossref. Use NCBI E-utilities rate limits and include tool/email parameters if running repeated searches.

4. Evaluate whether each paper supports the segment

Use a conservative support scale:

• strong support: the paper directly tests the same relationship/mechanism/method and the result supports the segment.
• partial support: the paper supports part of the segment, a related model, or a narrower condition.
• background support: the paper supports field context, not the specific claim.
• contradictory/limiting: the paper conflicts with or narrows the claim.
• metadata-only candidate: title/metadata suggest relevance, but abstract/full text has not been checked.

Never cite a metadata-only candidate as support without checking the abstract or publisher page. If a paper is a review, label it as review/context and avoid using it as primary evidence for an experimental claim when primary articles are available.

5. Export reference-manager file

Default behavior: write one reference-manager file; support publication time filters with --from-year and --to-year; for long or ambiguous texts, use --with-artifacts so the HTML browser is available.

Default file is references.enw (EndNote tagged export). Optionally references.ris (if the user requests RIS) or references.rdf (Zotero RDF). If the user asks to choose the download format, treat ENW, RIS, and Zotero RDF as the supported options and return only one export file unless they explicitly ask for multiple. Do not invent missing fields: if DOI, pages, volume, or issue are missing, leave them absent. See references/ris-endnote.md for format details.

6. Optional review artifacts

Generate review artifacts (HTML/TSV/JSON/report) for long or ambiguous runs — they are the primary way the user browses, filters, and selects candidates. Use --with-artifacts when the text is long, the query is broad, or the user needs manual curation. Report the HTML visualization path prominently when artifacts are enabled, and generate TSV/JSON/report alongside the HTML.

7. Report results

Unless the user asks for a different format, return:

交互式引用浏览器
- [absolute path to citation_visualization.html]  ← 在浏览器中打开此文件，可筛选/选择/下载引用

检索范围
- [Nature Portfolio / Science family / Cell Press / flagship only, plus date limits]

分段引用对应关系
S001: [source segment]
  - [Author, year, title, journal, DOI]
  - 支撑等级: [strong/partial/background/limiting/metadata-only]
  - 插入建议: [e.g. after sentence / after clause]

导出文件
- [absolute path to references.enw / references.ris / references.rdf]

风险和缺口
- [missing full-text check, contradictory evidence, no direct CNS literature, etc.]

Put the HTML browser path FIRST, above everything else, so the user can immediately open and browse candidates. If no suitable CNS/Nature-series paper exists, say so plainly and suggest the best nearby options from non-CNS literature only if the user wants broader coverage. If the text is long, mention the batch strategy used, especially when you limited the run with --batch-size or --max-segments.