Citation
Unified citation skill covering discovery, verification, ranking, BibTeX generation, and finalization. Finds citation gaps in LaTeX drafts, retrieves real candidate papers, verifies via academic APIs, ranks with a quality formula, and produces provenance-aware BibTeX. Use when user says "curate citations", "verify references", "find citations", "整理引用", "验证引用", or needs citation management for paper writing.From its SKILL.md
npx -y skills add jacazjx/AI-Research-Orchestrator --skill citationAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
8.0 KB, as published. Nobody here has run it
Citation
Overview
A unified citation management skill covering the full lifecycle: discovering gaps, searching for papers, verifying authenticity, ranking candidates, generating BibTeX, and producing a finalized citation index.
Purpose
- Find and fill citation gaps in LaTeX manuscripts
- Verify all paper citations are authentic and properly attributed
- Rank candidate papers using a repeatable quality formula
- Generate BibTeX with explicit provenance fields
- Replace preprints with formally published versions
- Produce a finalized citation index with verification status
Workflow
1. Detect Gaps
Scan documents for citation markers that indicate missing references:
- LaTeX markers:
[cite],[citation needed],\todo{cite} - Plain-language notes: "needs citation", "find reference for this"
- Chinese prompts: "需要引用", "找文献支持"
- Cross-reference
\cite{}commands againstreferences.bibentries to find uncited references and undefined keys
2. Generate Search Targets
For each citation gap:
- Reduce the paragraph to one verifiable statement
- Extract 3-8 search terms (task, method, population, metric, constraint)
- Generate English keywords even if the draft is in Chinese
- Note the type of evidence needed (benchmark result, survey, causal claim, system design, theory, dataset paper)
3. Search & Gather
Query academic APIs in priority order:
| Priority | Source | Use |
|---|---|---|
| 1 | Existing project .bib files and verification ledger | Reuse already-verified citations |
| 2 | User-level persistent paper library | Cross-project reuse |
| 3 | Semantic Scholar, OpenAlex | Candidate discovery |
| 4 | CrossRef, DOI resolver | DOI confirmation |
| 5 | DBLP | CS bibliographic records |
API endpoints:
| API | Endpoint |
|---|---|
| Semantic Scholar | api.semanticscholar.org/graph/v1/paper/search |
| CrossRef | api.crossref.org/works |
| arXiv | export.arxiv.org/api/query |
| DBLP | dblp.org/search/publ/api |
| OpenAlex | api.openalex.org/works |
For each candidate, collect: title, authors, year, venue, DOI, abstract, citation count, source URL.
Preprint handling: Treat arXiv/bioRxiv preprints as discovery hints. If a formally published version exists, cite that instead with its DOI. Keep preprints only when no formal version exists, and label them explicitly.
Semantic Scholar API key: Check --semantic-scholar-api-key, then SEMANTIC_SCHOLAR_API_KEY env var, then persisted local secret. Continue with free shared pool if no key is available. Never write API keys into repo files. Never block workflow solely because a key is missing.
4. Verify & Rank
For verification sources, grades, and workflow, see references/citation-standards.md (Citation Verification Methodology section).
Verification -- for each citation (existing and new):
- Existence: paper found in academic databases
- Metadata accuracy: authors, year, venue, title match
- DOI validity: DOI resolves correctly
- Content alignment: claim accurately reflects cited content
- Attribution: original sources cited (not just reviews)
Verification by type:
| Citation Type | Required Verification |
|---|---|
| Journal article | DOI verification via CrossRef |
| arXiv preprint | arXiv ID verification |
| Conference paper | DBLP/Semantic Scholar verification |
| Book | ISBN verification |
| Website | URL accessibility check |
Severity levels:
| Issue | Severity |
|---|---|
| Fabricated citation (not in any database) | CRITICAL |
| Wrong claims attributed to source | HIGH |
| Retracted source cited | HIGH |
| Preprint without version number | MEDIUM |
| Incomplete bibliographic info | LOW |
Ranking -- score candidates using the rubric from references/quality-scoring.md:
- Venue score (CCF A/B, JCR Q1/Q2, high impact factor preferred)
- Freshness score (last 5 years preferred)
- Citation count, impact factor, relevance, evidence type
- Publication bonus (peer-reviewed > preprint)
- Apply a minimum relevance gate. A lower-scoring paper that directly supports the claim can beat a higher-scoring but weakly related paper.
5. Output
BibTeX with provenance fields -- for each accepted citation:
@article{author2024method,
title = {Exact Title},
author = {Last, First and Last, First},
year = {2024},
journal = {Venue Name},
doi = {10.xxxx/yyyy},
url = {https://doi.org/10.xxxx/yyyy},
x-bib-source = {crossref},
x-bib-source-url = {https://api.crossref.org/works/10.xxxx/yyyy},
x-verified-with = {semantic-scholar, crossref},
x-verified-at = {2026-03-21},
x-quality-score = {85},
x-verification-status = {verified},
}
Do not fabricate missing BibTeX fields. If the source record is incomplete, note what is missing.
Citation index -- save to paper/citation-index.md with:
- Summary: total citations, verified count/percentage, needs-attention count, fabrication flags
- Verified citations table (key, title, year, venue, DOI, verified-by, quality score)
- Needs-verification table (key, issue, action needed)
- Fabrication flags table (key, issue, severity) -- HIGH PRIORITY
- Claim-citation mapping (well-supported claims and claims needing support)
- Timestamped verification log
Sync and persist: After every run, sync accepted citations into .citation-curator/verification-ledger.json and the user-level paper library. If a target .bib file was provided, sync the ledger with it. Keep user-edited BibTeX entries intact.
Helper Scripts
See scripts/ for implementation utilities:
| Script | Purpose |
|---|---|
scripts/extract_citation_needs.py | Find citation gaps in .tex files |
scripts/score_papers.py | Rank candidates by quality score |
scripts/fetch_verified_bibtex.py | End-to-end verified citation workflow |
Exception Handling
- Older papers: Allow when seminal, standards-defining, or required for classic methods. Label as
seminal exception. - Non-CCF/JCR venues: Allow for interdisciplinary fields when the paper is otherwise strongly supported.
- Preprint-only: If user explicitly allows and no formal version exists, keep outside the final verified bibliography or label as
fallback preprint. - Rate limits: Use long backoff for shared Semantic Scholar pool. Do not surface failed intermediate attempts in user-facing output.
Key Rules
- Do not invent papers, DOIs, venues, citation counts, or impact factors
- Use ONLY academic database APIs (not web search) for verification
- Every citation should be verified -- flag any potential fabrications as CRITICAL
- Do not keep preprints when a verified formal publication exists
- Do not claim authenticated verification when no valid API key was used
- Do not overwrite user-authored
.bibentries -- store extra state in the ledger - Do not guess CCF tier or JCR quartile -- mark as unknown when unverifiable
- Surface disagreements between DBLP, Crossref, Semantic Scholar, and publisher pages
- Cross-verify with multiple sources when possible
References
references/quality-scoring.md- Scoring formula and tie-break rulesreferences/source-verification.md- DOI checks, preprint replacement, BibTeX provenancereferences/citation-standards.md- Citation grading criteria and paper phase citation rules