citeclaw-bibliography-curation

name	citeclaw-bibliography-curation
version	1.0.1
description	Use CiteClaw to build, clean, expand, and verify bibliographies for surveys, papers, and literature reviews. Trigger this whenever the user asks to collect references, grow a bibliography to dozens or hundreds of papers, clean BibTeX, map literature by topic, turn arXiv/DOI/URLs into citations, or create a survey-style literature map. Use this especially for research writing workflows where CiteClaw alone is not enough and you need a robust process that combines CiteClaw, web search, and manual verification.

CiteClaw Bibliography Curation

Use this skill when the user is not just asking for a single citation, but for a real literature workflow:

• expanding a bibliography
• cleaning noisy BibTeX
• curating references for a survey
• building a topic-wise literature map
• finding representative papers for a theme
• reconciling arXiv IDs, DOIs, anthology pages, and URLs

This skill complements citeclaw-cli. citeclaw-cli gives you commands. This skill gives you the workflow and decision rules.

Primary CiteClaw scope

Use CiteClaw for these command families:

• citeclaw citoid / citeclaw citation
• citeclaw crossref
• citeclaw semantic-scholar
• citeclaw cite, cite-pdf, fetch-pdf, openurl-resolve, batch

Fast commands

Use these as canonical examples:

npx citeclaw citoid formats
npx citeclaw citoid bibtex "10.1021/acsomega.2c05310"
npx citeclaw crossref "10.1021/acsomega.2c05310" --json
npx citeclaw semantic-scholar "10.1021/acsomega.2c05310" --json

Advanced Semantic Scholar

npx citeclaw semantic-scholar api /paper/search --params '{"query":"transformer","limit":3}'
npx citeclaw semantic-scholar paper-search "transformer attention" --limit 5 --fields "title,year,authors,url"

CiteClaw guardrails

• Prefer --json for machine-consumable output.
• Respect API limits; use --s2-api-key for Semantic Scholar to avoid 429.
• For PDF and batch flows, keep --profile and logs for diagnosis.

What this skill is for

This skill is optimized for the messy reality of research writing:

• some references are clean DOI/arXiv inputs
• some are only known by title or GitHub list entries
• some APIs rate-limit
• some metadata sources return the wrong paper
• a paper may need a small main bibliography and a large supplemental bibliography

The goal is not merely to produce BibTeX. The goal is to produce a bibliography that is:

• accurate enough to compile cleanly
• structured enough to support a survey narrative
• large enough to cover the field without flooding the main text

Default workflow

Follow this sequence unless there is a strong reason not to.

1. Separate the task type first

Classify the user request into one of these modes:

• single-citation: one or a few citations, high precision
• main-bibliography: the core papers that will be cited in the main text
• supplemental-bibliography: broad coverage, often 50-200+ entries
• literature-map: references grouped by topic, benchmark, or method family

Use different standards for each mode:

• main-bibliography: prioritize precision and canonical papers
• supplemental-bibliography: prioritize breadth, deduplication, and grouping

2. Start with canonical identifiers

Prefer, in order:

1. DOI
2. arXiv abs URL
3. ACL/EMNLP/NAACL anthology URL
4. official project or official benchmark page
5. title search only as fallback

If the user gives a vague title or a list from a GitHub awesome repo, do not trust the list blindly. Resolve each candidate to a canonical identifier first.

3. Use CiteClaw where it is strongest

Use CiteClaw first for:

• DOI -> BibTeX
• arXiv -> BibTeX
• anthology URL -> BibTeX
• quick Crossref lookups
• quick Semantic Scholar lookups when the API is responsive

Typical commands:

npx citeclaw citoid bibtex "https://arxiv.org/abs/2305.19860"
npx citeclaw citoid bibtex "10.1145/3589334.3648158"
npx citeclaw citoid bibtex "https://aclanthology.org/2023.emnlp-main.398/"
npx citeclaw crossref "conversational recommender systems survey" --json

Use the npm-hosted npx citeclaw ... form above as the default documented interface. Only switch to a local checkout if you explicitly need to inspect or patch CiteClaw itself.

Runtime sync

Normal citation commands can run directly. Use explicit sync commands when you want broader translator coverage or local CSL rendering:

npx citeclaw translators sync
npx citeclaw styles sync

Use translators sync to pull additional official/community translators. Use styles sync when you need local cite-style rendering or a larger CSL style set.

When CiteClaw is not enough

Do not force everything through CiteClaw. Switch methods when needed.

Use web search when:

• title-only search through CiteClaw is unstable
• Semantic Scholar is rate-limited
• Crossref returns books, theses, or irrelevant near-matches
• you need recent official platform or regulator documents

Use arXiv page metadata when:

• arXiv API resolution is flaky
• curated lists only provide arXiv links
• you need a large batch of reliable title/author/year extraction

Reliable fallback pattern:

1. scrape or collect https://arxiv.org/abs/<id>
2. read citation_title, citation_author, and citation_date meta tags
3. generate @misc entries from these fields

This is often more robust for bulk survey expansion than trying to title-search APIs under rate limits.

Use curated lists strategically

Curated GitHub lists are useful for recall, not for truth.

Use them to:

• discover candidate arXiv IDs
• identify subtopics and communities
• bootstrap supplemental bibliographies

Do not use them as final metadata sources. Always resolve entries to canonical identifiers and regenerate BibTeX.

Main bibliography vs supplemental bibliography

For surveys and broad literature reviews, use two layers.

Main bibliography

Keep this small and argument-driven.

Include:

• canonical papers
• surveys
• benchmarks
• representative methods
• official policy/platform documents directly used in the main argument

In the main text, cite selectively and synthetically. A paragraph should usually cite:

• one canonical paper
• one contrasting or newer paper
• one benchmark or survey if needed

Avoid long undifferentiated citation dumps in the core narrative.

Supplemental bibliography

Put broad coverage here.

Use it for:

• topic-wise expansion to 50-200+ entries
• appendix literature maps
• “additional literature” sections grouped by theme

Typical split:

• references.bib: core papers used in main argument
• references_supplement.bib: broad coverage

This lets the paper look like a thoughtful survey instead of a citation dump while still achieving field coverage.

Literature-map workflow

When the user wants a survey or broad coverage, group papers by research function, not by chronology alone.

Useful groupings:

• surveys
• task-defining papers
• benchmarks
• method families
• safety / fairness / governance
• domain-specific applications

For each group:

1. gather candidate IDs
2. deduplicate
3. generate BibTeX
4. write a short synthesis sentence explaining what that group contributes

The synthesis is as important as the references. The bibliography should support an argument, not replace one.

Deduplication rules

Before adding anything to the final bibliography:

• deduplicate by DOI if available
• otherwise deduplicate by arXiv ID
• otherwise deduplicate by normalized title

If the same work appears as arXiv and conference:

• use the venue version when the paper is clearly published and metadata is clean
• otherwise use arXiv for speed and consistency

Be consistent within one paper. Do not mix random venue versions and arXiv preprints unless there is a clear reason.

Citation quality rules

Treat main-text citations as claims with burden of proof.

• For technical claims: prefer papers, benchmarks, or official docs
• For policy claims: prefer regulator or government pages
• For platform/product claims: prefer official company announcements
• For public discourse or panic framing: it is acceptable to use major media, but only as background, not as technical evidence

Do not let media reporting become the backbone of a scientific argument.

Large-scale expansion pattern

When the user asks for something like “expand this to ~200 references”:

1. keep the main bibliography stable
2. identify 3-6 high-yield curated sources
3. extract candidate arXiv IDs or DOIs
4. bulk-resolve metadata
5. create a supplemental .bib
6. generate an appendix literature map grouped by topic
7. compile and verify no missing citation keys

This is usually better than trying to inflate the main bibliography.

Verification checklist

Before finishing, always check:

• all cited keys exist
• there are no unresolved citations in the LaTeX log
• core main-text claims are backed by the right type of source
• bibliography growth has not turned main-text paragraphs into citation dumps
• appendix or supplemental grouping is coherent

Useful checks:

latexmk -pdf -interaction=nonstopmode main.tex
rg -n "undefined citations|undefined on input line|I didn't find a database entry" main.log main.blg

Output patterns

If the user wants a bibliography build, prefer outputs like:

• references.bib
• references_supplement.bib
• sections/appendix_litmap.tex

If the user wants a report rather than files, summarize:

• what sources were used
• how many entries were added
• how they were grouped
• what still needs manual review

Guardrails

• Do not trust one API result without sanity-checking the title.
• Do not treat a curated list entry as verified metadata.
• Do not let supplemental coverage dilute the main-text argument.
• Do not use weak title-only matches when an official or canonical identifier can still be found.
• Do not keep hammering rate-limited APIs; switch tools.

Example use cases

Example 1 User: “Use CiteClaw to clean this BibTeX and fix the broken arXiv references.” Action: use CiteClaw first, repair bad identifiers, regenerate clean entries, and verify compile.

Example 2 User: “Expand this survey bibliography to around 200 papers.” Action: keep core references stable, build a supplemental bibliography from curated lists plus canonical metadata, and generate an appendix literature map.

Example 3 User: “Find representative work on agentic recommendation, shopping benchmarks, and agent safety.” Action: build grouped paper pools by theme, select representative papers for the main text, and place the long tail into supplemental references.