PubMed Citation Batch Verifier

The problem it addresses

Medical documents often contain citations that fail in two different ways: the PMID does not exist, or the PMID exists but points to a different paper. The second failure mode is more subtle, especially in AI-generated reviews, textbooks, or knowledge-base entries, because the citation can still look valid while failing to support the claim it is attached to. This skill is designed for batch citation checking: it extracts PMID values from documents, then compares them against nearby citation metadata such as title, authors, journal, and year.

Core capabilities and workflow

The verifier assigns each citation one of five states:

• Correct: the PMID exists and matches the claimed paper metadata.
• Mismatched: the PMID exists but points to a different paper.
• Partially matched: some metadata align, but title or author details still differ.
• Invalid: the PMID cannot be found in PubMed.
• Needs confirmation: the source document does not provide enough citation metadata to compare.

During processing, it scans files such as .html, .md, .txt, and .json for PMID: patterns or PubMed URLs, then parses surrounding text to extract authors, title, journal, and year when available. It fetches real metadata from the PubMed esummary API and compares that against the claimed citation. The matching logic combines token-level overlap with sequence similarity, which helps with abbreviations, word-order changes, and minor spelling differences. Cached PMID lookups can speed up repeated validation runs.

Optional features include Crossref DOI verification and candidate PMID suggestions for mismatched citations. Results can be exported as an HTML report, JSON, or plain terminal text, making them easier to review or feed into downstream pipelines.

Boundaries and caveats

This skill fits systematic reviews, AI-generated medical content, pre-submission citation checks, medical knowledge bases, and pharmacovigilance documentation. However, it depends on whether the document contains enough citation metadata. If only a bare PMID is present, the result may remain uncertain. Fuzzy matching can also produce borderline cases that still require human judgment. It is not a full fact-checking tool and does not replace reading the paper itself, checking conclusions, or reviewing how the citation is used in context.