菜单
Use when citations, quotations, page numbers, bibliography entries, or source-supported claims must be checked for existence, metadata accuracy, textual fidelity, and claim–source fit.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when a manuscript contains recurring concepts, translations, abbreviations, proper names, or near-synonyms whose meanings and forms must remain stable across the argument.
Use when a humanities researcher needs to transform verified sources into a structured scholarly conversation rather than a chronological list of summaries.
Use when a humanities researcher has a broad topic, intuition, corpus, or problem but does not yet have a precise, contestable, and researchable question.
Use when a humanities argument depends on a concept whose history, competing definitions, translations, or boundaries must be reconstructed before the concept can support new claims.
Use when an interpretation must be grounded in specific textual, visual, rhetorical, formal, interface, archival, performative, or material features rather than thematic summary or theoretical labeling.
Use when a research question, scope boundary, and source base exist but the manuscript still needs a defensible sequence of claims, warrants, evidence, objections, and section functions.
| name | auditing-citations |
| description | Use when citations, quotations, page numbers, bibliography entries, or source-supported claims must be checked for existence, metadata accuracy, textual fidelity, and claim–source fit. |
| version | 1.0.0 |
| language | en |
| license | MIT |
Audit the evidentiary chain between a manuscript and its sources. This skill does not beautify a bibliography or infer missing publication details. It creates a traceable record showing what was checked, against which source, with what result, and what remains unresolved.
Citation accuracy has at least two distinct dimensions. Bibliographic verification asks whether the cited object exists and whether author, title, date, edition, journal, volume, issue, pages, DOI, archival identifier, and translator information are accurate. Claim–source verification asks whether the cited passage actually supports the sentence, interpretation, historical statement, or quotation to which it is attached. A record can pass the first test and fail the second.
Accepts a manuscript, notes, bibliography, citation export, quotation list, source files, authoritative catalogue records, or a prior CitationRecord set.
Requires the text containing the citations and enough source access to distinguish checked items from inaccessible ones. Direct quotations require the exact edition or a clearly identified equivalent edition.
Produces normalized CitationRecord objects, an itemized audit table, discrepancy notes, submission-blocking issues, and a citation gate report.
May produce a corrected bibliography proposal, a source-access request list, duplicate-record map, or a list of claims that need weaker wording or different evidence.
Fails when the cited source cannot be identified, a quotation cannot be located, an edition or pagination conflict prevents reliable matching, or the requester asks the agent to invent or conceal missing details.
Guarantees that every audited item receives an explicit status and that inaccessible or mismatched items are not silently represented as verified.
Does not guarantee that inaccessible sources are accurate, that the bibliography is exhaustive, that a verified source makes the manuscript's interpretation correct, or that a journal will accept the citation style.
Use this skill before submission, after AI-assisted drafting, after importing references from a manager, when direct quotations have been translated or paraphrased, when a reviewer questions a source, or whenever references were assembled from secondary mentions rather than primary consultation.
Do not wait until the final day when the manuscript contains many direct quotations. Citation auditing is most reliable when repeated after major drafting rounds and again at the submission gate.
At minimum provide:
When any input is missing, record it as a limitation. Do not treat a citation string as evidence that the underlying work was consulted.
Extract every in-text citation, footnote, endnote, direct quotation, epigraph, translated passage, data attribution, archival reference, URL, DOI, and bibliography entry. Assign a stable CitationRecord ID. Reconcile obvious duplicates without deleting provenance.
Label what each citation is being asked to do:
A source used only as background MUST NOT be reported as direct proof of a contested claim.
Check author or editor, title, container title, publication year, edition, translator, volume, issue, page range, publisher, DOI, URL, archive, collection, and item number as relevant. Prefer the source itself, publisher metadata, DOI registry, library catalogue, or archival finding aid over copied citations from unrelated webpages.
Record discrepancies exactly. If two legitimate editions differ, preserve both identities and identify the edition used in the manuscript.
Compare wording, spelling, punctuation, ellipses, brackets, emphasis, and surrounding context. Confirm page, paragraph, section, timestamp, image number, folio, or stable location. For translations, distinguish:
Do not silently modernize or improve quoted language.
Read enough surrounding context to determine whether the source supports the attached claim. Classify fit as:
A shared keyword is not sufficient support. A citation to a theorist's name does not validate every later use of that theorist.
Determine whether a citation applies to one sentence, several sentences, or a whole paragraph. Flag ambiguous citation placement, citation clusters where no source is mapped to a specific claim, and paragraph-final citations that appear to support more than they actually do.
Identify cited-but-unlisted works, listed-but-uncited works, duplicate identities, inconsistent author names, year suffix errors, broken cross-references, missing translators, and mismatched titles. Formatting corrections come after identity is secure.
Use only explicit states:
verified-primaryverified-secondaryresearcher-suppliedinferenceunverifiedunknowndisputedDo not use “verified” when only a search snippet or another bibliography was checked.
Submission-blocking issues include fabricated or absent sources, unverifiable direct quotations, materially wrong page numbers, citations that contradict the attached claim, and unresolved identity conflicts affecting the argument. Minor punctuation differences in a citation style are normally non-blocking.
Return PASS only when every citation has a status and no blocking item remains. Use CONDITIONAL PASS when non-blocking format issues or clearly listed access limitations remain. Return FAIL when any central quotation, evidentiary citation, or bibliographic identity is unverified or contradicted.
Stop and request researcher action when:
Do not “repair” these problems through guessing.
The skill is complete only when:
Never invent an author, title, page, DOI, issue number, archive identifier, quotation, or source content. Never use search-result snippets as substitutes for reading the relevant passage. Never convert a likely match into a verified identity. Never hide a failed citation by deleting it without reporting the affected claim. Never translate a quotation and present it as an established published translation unless that translation was checked.
When source access is partial, say exactly what was checked. “Metadata verified; passage not checked” is preferable to a polished but false assurance.
Use templates/citation-audit.md and include:
Gate: Citation Audit
Status: PASS | CONDITIONAL PASS | FAIL
Evidence checked:
Blocking issues:
Non-blocking risks:
Researcher decisions required:
Required next action:
Next skill:
Audit all citations and direct quotations in this manuscript against the supplied PDFs. Distinguish metadata verification from claim–source fit and fail the gate if a central quotation cannot be checked.
Make the references look complete, supply plausible page numbers, and do not tell the reviewer which sources I could not access.
The bad invocation requests concealment and fabrication. The skill MUST refuse it and preserve the unresolved state.
checking-terminology-consistency when translated names or concepts vary.responding-to-peer-review when the audit addresses reviewer concerns.verifying-before-submission after all blocking citation issues are resolved.structuring-humanities-argument when a claim must be narrowed after source mismatch.An audit is bounded by source access, edition identity, language competence, and the quality of authoritative metadata. It can establish that a passage exists and is represented accurately; it cannot decide by itself whether a complex interpretation is persuasive. Specialist review may be required for rare languages, manuscripts, archival conventions, legal restrictions, or contested editions.