paper-citation-format

name

paper-citation-format

description

Add, search, normalize, and format references for academic papers, theses, and literature reviews. Use when an AI assistant needs to search for supporting papers, read article content or abstracts, insert missing citations, map claims to provided or discovered sources, rewrite in-text citations, generate or clean a bibliography, compare multiple AI citation drafts, convert between APA, IEEE, Chicago, MLA, GB/T 7714, BibTeX, or journal-specific styles, or route the task through a reusable journal-family citation profile.

Paper Citation Format

Overview

Use this skill to keep citation work consistent across two layers:

1. An AI-provider layer that keeps prompts portable across Codex, ChatGPT, Claude, Gemini, DeepSeek, Qwen, or similar assistants
2. A journal or style layer that chooses the right citation profile before editing

Treat this skill as the citation policy layer, not as a single-model prompt. The same source inventory, validation rules, and no-fabrication policy should survive even when the task moves between different AI systems.

The most important operating mode is active source discovery: if the user gives a manuscript passage but not enough references, search for relevant papers first, verify what each source actually supports, and only then annotate the passage with citations.

When output-format fidelity matters, prefer the local assets under assets/official-citation-assets/ so the skill can reference real CSL files or official family bibliography packages instead of relying on prose rules alone.

Do not invent authors, years, DOIs, page ranges, issue numbers, conference names, journal names, URLs, or publisher metadata. If the metadata is incomplete, flag the gap and keep the uncertainty visible.

When To Use

Use this skill for requests such as:

• "Add citations to this related-work section"
• "Convert this bibliography from IEEE to APA"
• "Make this reference list match the journal style"
• "Search papers for these claims and cite them"
• "Read article content and annotate each sentence with references"
• "Use the same citation task across ChatGPT, Claude, and Gemini"
• "Prepare citations for a Chinese paper using GB/T 7714"
• "Clean references for a thesis or journal manuscript"
• "Build a reusable prompt packet for citation generation across multiple AI tools"

Working Method

1. Lock the target task before editing:
   • Add missing citations
   • Search for supporting papers and then add citations
   • Convert citation style
   • Repair in-text citations
   • Clean the bibliography only
   • Produce BibTeX, CSL JSON, or another machine-readable artifact
2. Lock the citation target:
   • In-text mode: numeric, author-date, notes, or footnotes
   • Reference list mode: appearance-ordered, alphabetical, or template-specific
   • Final output: revised prose, bibliography, BibTeX, table, or another requested format
3. Route the AI layer:
   • Open references/ai-routing.md for provider-neutral task packaging
   • Open references/multi-ai-prompts.md only when you need a reusable prompt wrapper for a specific AI assistant
   • Open references/source-discovery.md when the task requires the model to search literature on its own and annotate claims from discovered papers
   • Open references/domain-search-templates.md when you need field-specific search-query patterns
4. Route the journal or style layer:
   • Open references/journal-profiles.md for the folder model
   • Open references/journals/catalog.md to choose a family baseline
   • If an exact journal profile exists under references/journals/<family>/<journal>/, prefer it over the family baseline
   • If no exact journal is modeled yet, use the nearest family baseline
   • If the user provides a local author guide or venue instructions that conflict with the library, use those instructions and fall back to custom-journal/profile.md
5. If the user has not supplied enough sources, run active source discovery:
   • Split the text into claims that require evidence
   • Search for candidate papers using claim keywords, domain terms, and likely venue vocabulary
   • Adapt queries to the field using references/domain-search-templates.md when the topic is specialized
   • Prefer primary papers, peer-reviewed sources, and official publisher pages when available
   • Open the paper landing page, abstract, or full text before citing it
   • Record what the source actually supports and what remains uncertain
6. Build a normalized source table before rewriting:
   • Author list as published
   • Publication year
   • Title
   • Container title
   • Volume, issue, pages when available
   • DOI, URL, arXiv ID, ISBN, or publisher page when available
   • Source type such as journal article, conference paper, thesis, dataset, software, standard, patent, or webpage
7. Map claims to sources conservatively:
   • Attach citations only to claims actually supported by the provided or verified source
   • Prefer primary sources for technical claims
   • Avoid citation padding
   • If a source was only discovered from a title list or abstract, do not overclaim support beyond what was actually checked
8. Rewrite the in-text citations and reference list together so they remain aligned
9. Finish with a consistency sweep:
   • Every in-text citation resolves to one bibliography entry
   • Every bibliography entry matches the selected profile
   • Duplicates are merged
   • Missing metadata stays visible

Editing Rules

• Preserve the manuscript wording unless citation work requires a local rewrite
• Normalize mixed citation styles into one system unless the user explicitly asks to preserve the mixture
• Keep author order exactly as published
• Treat DOI, URL, and identifier formatting consistently across the whole list
• Preserve language-specific details when the profile requires bilingual or transliterated entries
• If the user asks for a polished draft despite missing metadata, mark placeholders clearly instead of guessing
• When citations are discovered by search, keep a brief evidence note for each newly added source so the mapping from claim to paper can be checked
• Do not cite a paper merely because its title sounds relevant; inspect the abstract, landing page, or full text first

Multi-AI Rules

• Keep one shared task packet across all AI providers
• Use thin provider wrappers instead of rewriting the whole citation policy for each assistant
• Preserve the same output schema so results can be compared side by side
• If a provider can browse, let it search and verify sources directly
• If a provider cannot browse, have it produce a search packet: query list, target claims, candidate papers to verify, and unresolved evidence gaps
• Compare cross-provider outputs in this order:
  1. Metadata accuracy
  2. Citation-to-bibliography alignment
  3. Match between the cited sentence and what the checked article content actually supports
  4. Match to the requested journal or style profile
  5. Handling of uncertainty
  6. Formatting polish

If two providers disagree, prefer the version better grounded in the supplied source metadata.

Output Rules

• If editing prose, return the revised passage and the matching bibliography entries
• If performing active source discovery, also return a short evidence table or search log for newly introduced sources
• If cleaning a full manuscript, preserve structure and limit edits to citation-related text unless asked otherwise
• If generating references only, return a clean list in the requested profile
• If generating BibTeX or another structured format, preserve machine-readable fields and flag missing required fields separately
• End with a short verification list when any metadata remains unresolved

Reference Files

Open references/ai-routing.md when the task must stay reusable across different AI assistants.

Open references/multi-ai-prompts.md when you need ready-to-send wrapper prompts for Codex, ChatGPT, Claude, Gemini, DeepSeek, or similar models.

Open references/source-discovery.md when the model must search for papers on its own, inspect article content, and annotate claims with newly found citations.

Open references/domain-search-templates.md when the search task needs domain-specific query design, venue keywords, or evidence priorities.

Open references/journal-profiles.md for the citation-profile folder model, then use references/journals/catalog.md to select the right family baseline.

Open references/citation-styles.md only for a concise reminder of common style mechanics after the target profile has already been chosen.

Inspect assets/official-citation-assets/ when the task needs local CSL assets, bibliography packages, or a manifest of refreshable citation-format sources.