メニュー
Verify citations in AI-generated or human-written text against the local knowledge base and external APIs. Catches hallucinated references, wrong metadata, and missing papers. Use when the user wants to check if citations are real and accurate.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | citation-check |
| description | Verify citations in AI-generated or human-written text against the local knowledge base and external APIs. Catches hallucinated references, wrong metadata, and missing papers. Use when the user wants to check if citations are real and accurate. |
Check whether citations in text are real and accurate — prevents AI-hallucinated references and metadata errors.
Important context: Roughly 40% of citations in AI-generated academic text may be hallucinated (fabricated papers, misattributed authorship, assembled metadata). Even human-written text frequently contains year/journal errors. The goal of this skill is to eliminate all citation problems before submission.
The user provides text to check (pasted inline, a file path, or a draft file in a workspace). If a workspace is available, validation is preferentially scoped to it.
Extract all citations from the text, recognizing these formats:
(Author, Year) / Author (Year) — parenthetical citations\cite{key} / \citep{key} / \citet{key} — LaTeX citations[N] — numbered citations (require a reference list)Apply a three-layer check to every citation:
Layer 1 — Local library match
autor search-author "<Author>" --top 5
autor search "<keywords from title>" --top 5
After finding a match in the local library, verify: author names, year, title, and journal are consistent.
Layer 2 — DOI / metadata cross-check If a match exists in the local library, read the DOI and detailed metadata from meta.json for cross-validation. If no local match is found, flag the citation — it is not in the workspace/knowledge base.
Layer 3 — Content consistency For key citations (those supporting core claims), load L2–L3:
autor show <dir_name> --level 3
Verify: does the text's description of the paper match the paper's actual content? Is there over-interpretation or selective quotation? If L3 is inferred_synthesis, use it as a triage signal and verify high-stakes claims against L4/source text.
Generate a verification report with a status label for every citation:
| Status | Meaning |
|---|---|
| VERIFIED | Found in local library with consistent metadata |
| METADATA MISMATCH | Paper found but author/year/title differs |
| NOT IN LIBRARY | Paper not present in the local library |
| CONTENT MISMATCH | Paper content does not match the in-text description |
| SUSPICIOUS | Cannot be verified; likely a hallucinated citation |
Provide specific correction suggestions for each problematic citation.
User says: "Check whether the citations in this passage are correct." → Extract citations, search for each in the local library, output verification report
User says: "Check citations in workspace/my-paper/introduction.md." → Read the file, extract citations, verify within workspace scope
Adversarial but evidence-bound Reviewer
Plan a literature review before drafting. Build a lean evidence-bound, taste-gated package using references.bib, reference-map.json, review-plan.md, evidence-ledger.md, and table-figure-plan.md.
Draft a fact-grounded, taste-gated review manuscript from an approved lean planning package. Uses references.bib, reference-map.json, review-plan.md, evidence-ledger.md, and table-figure-plan.md; use only after plan is approved.
Understand the autor codebase at a high level. Use when the user asks how to use autor, what built-in skills it has, what features it supports, which workflow to choose, or wants a project overview before using other skills.
Write an exploratory literature review based on papers in an autor workspace. For approved review-article workflows, respect the canonical references.bib / reference-map.json / evidence-ledger contract produced by plan.
High-level orchestrator for iterative review planning and hypothesis evolution. Builds the current autor planning package, triggers targeted acquisition, maintains corpus layers, and stops before writing.