| name | literature-review-search |
| description | Use when finding references, building literature review tables, identifying up-to-date or highly cited papers, comparing research methods, verifying citation quality, mapping references to claims, or converting AI literature notes into SCI-ready bibliography inputs |
Literature Review Search
Purpose
Use this skill to turn a research topic into a citation-safe literature review table and a claim-evidence map. The default output language is Chinese unless the user asks otherwise.
Core Workflow
- Define the research scope:
- problem domain;
- method keywords;
- target application;
- preferred venues or source types;
- time window and "classic/high-citation" exceptions.
- Search online when references, publication status, citation recency, journal names, or latest papers are requested.
- Prefer primary sources:
- publisher pages, DOI pages, arXiv, IEEE, ACM, Elsevier, Springer, CVF, MDPI, Nature, NASA/ESA/CCSDS pages;
- use secondary summaries only as leads.
- Verify each candidate:
- exact title;
- venue or institution;
- publication year;
- DOI/arXiv/publisher URL where available;
- method and conclusion from abstract/full text, not from another model's paraphrase.
- Classify evidence quality:
- A. 可直接引用: title, venue, year, and source are verified.
- B. 可候选引用: relevant but missing DOI/full metadata or only preprint available.
- C. 仅作线索: found in model output or blog/source list but not yet verified.
- D. 不建议引用: exact title cannot be verified or source appears hallucinated.
- Produce the requested table. If the user only asks for the four-column format, use exactly:
- 参考文献名称;
- 研究方法和研究结论;
- 期刊/会议/机构名称;
- 发表时间.
- For paper-building tasks, add optional columns only when useful:
- 与本研究的关系;
- 可支撑的论文 claim;
- 核验状态;
- 来源链接.
- Never convert unverified model-generated references into final citations.
Search Strategy
Build several query groups instead of one broad query:
- core topic query: problem + application;
- method query: key algorithm or system method;
- benchmark query: dataset / evaluation / baseline;
- mission or industry query: NASA, ESA, CCSDS, mission names;
- review query: survey, review, state of the art;
- exact-title query for suspicious candidates.
For "up-to-date, highly cited" requests:
- include recent papers from the last 3-5 years;
- include older classics when they define the field;
- do not invent citation counts;
- if citation counts are not verified from a citation index, say "高被引待核" or omit the count.
Output Discipline
- Separate literature-supported facts from project claims that still need experiments.
- Use cautious language for arXiv papers, workshops, and technical reports.
- Do not claim novelty from absence of evidence after a shallow search.
- Do not say a method is "first", "best", or "state of the art" unless the source itself and current literature support it.
- Preserve links to sources used.
- Flag questionable references explicitly instead of silently dropping them when they came from the user.
Useful References
Read these bundled files when needed:
references/search_prompt.cn.md: reusable Chinese prompt adapted from the research-guide PDF.
references/output_contract.cn.md: table formats, vetting labels, and claim-evidence mapping rules.