CNKI Literature Search and Word Citation Annotation

Problem

Writing a paper in a Chinese academic journal often fails less because of one bad source and more because the claim-to-evidence-to-Word workflow breaks down. Teams may search by paper title or abstract keywords, then discover that the year window is exceeded, a repost is treated as a journal article, download counts are mistaken for citation counts, and core-journal status is never verified. Later, Word citation formatting can still be inconsistent: numbers are not continuous, superscript formatting leaks into punctuation, reference entries keep DOI: or CNKI record numbers, and citations cluster in one section.

How It Works

The skill treats the manuscript as evidence slots. It starts from AGENTS.md, the full paper, and the requested years, counts, topics, and annotation style, with a default window of the current Gregorian year and the two preceding years. If a fixed project year conflicts with that rolling window, the conflict must be surfaced rather than claiming the last-three-years requirement is met. Each target sentence is then decomposed into research object, mechanism, scope, and evidence type, avoiding title-only or keyword-only search.

• For CNKI access, it prefers an already signed-in Chrome session and uses chrome:control-chrome; if real-time pages cannot be checked, results must be marked as unverified for citation counts.
• Each candidate is checked against year, formal journal type, bibliographic completeness, direct support, and extrapolation risk; claims about causality, effectiveness, or digital governance also receive counter-evidence searches for limitations, failures, false positives, risks, and conditions.
• For Word edits, it uses documents:documents for minimal changes: the citation number sits in its own run as true superscript, while the following punctuation is placed in a separate run and returned to normal baseline.
• Before delivery, scripts/audit_docx_citations.py validates first-occurrence order, continuous numbering from [1] to [N], chapter-level citation distribution, and cleanup of reference entries.

Boundaries

The scope is explicit: formal journal articles take priority, while dissertations, conference papers, newspapers, yearbooks, and encyclopedias are excluded by default; laws, policies, and official statistics still cite the issuing authority's original text. It does not fabricate DOIs, page numbers, issue numbers, or core-journal markers, and it does not move citations to sentences they cannot support just to make the distribution look even.