菜单
**Citation Checker & Verification Agent**: Deploys 5 parallel agents to audit every citation — cross-references, APA 7 formatting, DOI validation, claim grounding, and completeness. Detects fabricated references. Produces a structured verification report.
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
**Systematic Literature Review Pipeline**: End-to-end systematic review — from research question through search strategy, screening, data extraction, quality assessment, synthesis, and PRISMA 2020-compliant manuscript. Produces .docx manuscript and .xlsx review matrix.
**Academic Editor Team**: Multi-pass editorial review — five sequential editors: (1) Structural Editor for argument flow, (2) Copy Editor for clarity, (3) Voice Editor for register consistency, (4) Citation Editor for reference verification, (5) Rubric Editor for grade prediction. Produces a unified editorial report.
**Academic Essay & Literature Review Writer**: Creates structured academic essays, systematic literature reviews, scoping reviews, and research papers with PRISMA compliance, PICO frameworks, and APA 7 citations. Outputs professional .docx files.
**Citation Fixer & Correction Agent**: Companion to citation-checker. Deploys 3 parallel agents to correct reference metadata, fix APA 7 formatting, and insert [CITATION NEEDED] markers. Creates backup before editing. Produces a changelog.
**Zotero Citation Helper**: Searches a Zotero library (CSL-JSON) to find references by author, title, keyword, or year. Returns citation keys [@citekey], formats APA 7 entries, generates reference lists, and suggests citations for claims.
**Academic Project Report Writer**: Creates structured project reports with business model canvases, concept notes, and theory-integrated sections. Professional .docx output with APA 7 citations, tables, and formatting.
| name | citation-checker |
| description | **Citation Checker & Verification Agent**: Deploys 5 parallel agents to audit every citation — cross-references, APA 7 formatting, DOI validation, claim grounding, and completeness. Detects fabricated references. Produces a structured verification report. |
| argument-hint | "file-path or paste document" [style: apa7] |
| user-invokable | true |
This skill deploys parallel agents to systematically verify every citation in an academic document. It catches fabricated references, misattributed claims, formatting errors, and missing sources — the most common and most damaging errors in academic writing.
Use when the user asks you to:
The skill needs at minimum:
.md, .txt, .pdf) or pasted textIf the document has no reference list (e.g., an early draft), the skill switches to reduced-scope mode: only Agent 5 (Completeness) runs, flagging uncited claims. Agents 1-4 are skipped since they require a reference list to work against. Tell the user: "No reference list found. Running completeness check only — add a reference list and re-run for full verification."
If the input is a .docx file, convert it to markdown first using Pandoc (pandoc input.docx -t markdown -o input.md) before parsing. If Pandoc is not available, ask the user to export as .md or .txt.
Optional:
Agents that verify references or suggest sources follow a 4-tier lookup chain, checking the researcher's own sources before web searching:
0. ZOTERO: Search the Zotero CSL-JSON library configured in the plugin's `.local.md` file (default: `references/salsu-library.json`) by author, title, year
├── Match found → Use structured metadata (authors, DOI, volume, pages) as ground truth
└── No match ↓
1. LOCAL PDFs: Search the local PDFs directory configured in `.local.md` by author surname, title keywords, or topic
├── Match found → Read PDF to verify content or extract metadata
└── No match ↓
2. NOTEBOOKLM: Query researcher's NotebookLM notebooks (if available)
├── Match found → Use notebook content for verification
└── No match or not configured ↓
3. WEB SEARCH: Search via WebSearch/WebFetch (fallback)
Graceful degradation: If any tier is unavailable (Zotero library not configured, no local PDFs directory, NotebookLM not set up), skip it silently and proceed to the next tier. The skill works with web search alone — local sources improve quality but are not required.
Why this order:
.local.md provide this. Also useful when a reference isn't in Zotero but the researcher has the paper from course materials.Zotero lookup: Read and parse the Zotero CSL-JSON library configured in the plugin's .local.md file (default: references/salsu-library.json). Search by author family name, title keywords, or issued year. Use the citation-helper skill's search logic (Operation 1). If the reference is found in Zotero, its CSL-JSON metadata is authoritative — use it to verify or correct authors, DOI, volume, issue, pages.
Local source matching: Search the local PDFs directory configured in .local.md using Glob for PDF filenames containing the author surname or title keywords (e.g., **/*Silvius*Schipper*.pdf, **/*complexity*leadership*.pdf). See references/verification-patterns.md for matching patterns.
NotebookLM integration: If the researcher has set up NotebookLM notebooks with collected sources, query them via /notebooklm before falling back to web search. This is optional — if no notebooks are configured, skip gracefully.
The skill deploys 5 parallel verification agents, each handling one class of checks. All agents run concurrently via the Agent tool and report back to a coordinator that merges findings into a single report.
┌─────────────────────┐
│ COORDINATOR │
│ (this skill) │
└────────┬────────────┘
│ launches 5 agents in parallel
┌───────────┬───────┼───────┬────────────┐
▼ ▼ ▼ ▼ ▼
┌─────────┐ ┌─────────┐ ┌────┐ ┌──────┐ ┌──────────┐
│ Agent 1 │ │ Agent 2 │ │ A3 │ │ A4 │ │ Agent 5 │
│ Cross- │ │ Format │ │DOI │ │Claim │ │Complete- │
│ Ref │ │ Check │ │Val │ │Ground│ │ ness │
└─────────┘ └─────────┘ └────┘ └──────┘ └──────────┘
Purpose: Ensure every in-text citation has a matching reference list entry, and every reference list entry is cited at least once.
Checks:
(Author, Year), (Author & Author, Year), (Author et al., Year), narrative forms like Author (Year)Output fields per issue:
| Field | Description |
|---|---|
| Location | Page/paragraph/line where the issue appears |
| Citation | The problematic citation text |
| Issue | What's wrong |
| Severity | Error / Warning |
| Suggestion | How to fix it |
Purpose: Verify that all citations and references follow APA 7 formatting rules.
In-text citation checks:
& (parenthetical) vs. and (narrative)et al. used correctly for 3+ authors from first citation(Organisation [ABBR], Year), subsequent (ABBR, Year)Reference list checks:
Surname, A. A.https://doi.org/ hyperlink (not doi: or dx.doi.org)(2nd ed.) not (Second edition)In before editor names in book chapters(Ed.) / (Eds.) formattingOutput fields per issue:
| Field | Description |
|---|---|
| Location | Position in document |
| Current | What's currently written |
| Issue | Which APA 7 rule is violated |
| Severity | Error / Warning / Style |
| Fix | Corrected version |
Purpose: Verify that references point to real publications. This is the primary defence against hallucinated references.
Checks:
10.XXXX/YYYY)Validation approach (4-tier):
Tier 0 — Zotero library (fastest, structured metadata):
.local.md file (default: references/salsu-library.json) for the reference by author family name and issued yearTier 1 — Local PDFs:
5. Search the local PDFs directory configured in .local.md for PDFs matching the reference: Glob **/*AuthorSurname*.pdf and Glob **/*TitleKeyword*.pdf
6. If found, read the PDF's first pages to extract full citation metadata
7. Cross-check against the reference entry
Tier 2 — NotebookLM (if available): 8. Query NotebookLM for the reference author/title to check if the researcher has collected it
Tier 3 — Web search (fallback):
9. For each reference NOT found in Zotero, local PDFs, or NotebookLM, search: "Author Surname" "Article Title" "Year"
10. Cross-check the returned results against the reference details
11. If DOI present, search: doi:10.XXXX/YYYY to confirm it resolves
Rate confidence: Verified (Zotero) / Verified (Local) / Verified (Web) / Likely Valid / Uncertain / Likely Fabricated / Cannot Verify
Output fields per reference:
| Field | Description |
|---|---|
| Reference | The full reference as written |
| DOI Status | Valid / Invalid / Missing / Malformed |
| Publication Status | Verified / Likely Valid / Uncertain / Likely Fabricated |
| Details Match | Which fields match/don't match the real publication |
| Notes | Any discrepancies found |
Purpose: Verify that claims in the text are actually supported by the cited source. This catches misattributions and misrepresentations.
Checks:
Verification approach (4-tier):
Tier 0 — Zotero abstract (quick grounding check):
.local.md file (default: references/salsu-library.json) for the cited source by author and yearabstract field is present, check the claim against the abstract textTier 1 — Local PDFs (full-text verification, highest confidence):
4. Search the local PDFs directory configured in .local.md for the cited source's PDF using Glob (author surname, title keywords)
5. If found, READ the actual PDF (use Read tool with pages parameter for large files) and verify the claim against the source text
6. This upgrades assessments from "Plausible" to definitively "Grounded" or "Unsupported"
7. Note the local path in the output for the user's reference
Tier 2 — NotebookLM (if available): 8. Query the researcher's NotebookLM notebooks for the cited source to access annotated content
Tier 3 — Title-based assessment (fallback): 9. If the source is not available through any local channel, assess based on title, journal, and known content
Confidence levels:
Output fields per flagged claim:
| Field | Description |
|---|---|
| Location | Where in the document |
| Claim | The statement being made |
| Cited Source | The reference being used to support it |
| Assessment | Grounded / Plausible / Questionable / Unsupported |
| Reason | Why the assessment was made |
| Suggestion | Possible correction or alternative source |
Purpose: Find statements that should have citations but don't, and assess overall citation coverage.
Checks:
What does NOT need a citation (don't flag these):
Output fields per issue:
| Field | Description |
|---|---|
| Location | Paragraph/section |
| Statement | The uncited claim |
| Issue | Missing citation / Over-reliance on single source / Citation desert |
| Severity | Error / Warning |
| Suggestion | What type of source would be needed |
| Zotero matches | Relevant references found in the Zotero CSL-JSON library with [@citekey] (if any) |
| Local sources | Relevant papers found in the local PDFs directory configured in .local.md (if any) |
Source suggestions follow the 4-tier lookup: First search the Zotero CSL-JSON library configured in the plugin's .local.md file (default: references/salsu-library.json) for relevant sources by topic keywords, title, or abstract — these come with ready-to-use [@citekey] citation keys. Then search the local PDFs directory configured in .local.md for local PDFs. Then check NotebookLM, then web search. Label suggestions with [ZOTERO], [LOCAL], or [WEB]. Prioritise sources the researcher already has.
Read the input document and extract:
Present a summary to the user:
Found X in-text citations across Y pages/sections, with Z entries in the reference list. Deploying 5 verification agents...
Launch all 5 agents in parallel using the Agent tool. Each agent receives:
Merge all agent results into a structured verification report using the merging rules below.
Each issue type is owned by exactly one agent. If two agents flag the same item, the owner's finding takes precedence and the duplicate is dropped.
| Issue Type | Owned By | NOT owned by |
|---|---|---|
| Ghost citations (cited but not in reference list) | Agent 1 | Agent 5 |
| Orphan references (in reference list but never cited) | Agent 1 | Agent 5 |
| Author name inconsistencies (in-text vs reference list) | Agent 1 | — |
| APA 7 formatting errors (in-text and reference list) | Agent 2 | — |
DOI format errors (doi: vs https://doi.org/) | Agent 2 | Agent 3 |
DOI structural validity (10.XXXX/YYYY pattern) | Agent 3 | Agent 2 |
| Publication existence verification | Agent 3 | — |
| Claim-source grounding | Agent 4 | — |
| Missing citations (uncited claims) | Agent 5 | — |
| Citation deserts and over-reliance | Agent 5 | — |
All agent findings map to a 3-tier report severity:
| Report tier | Agent 1 | Agent 2 | Agent 3 | Agent 4 | Agent 5 |
|---|---|---|---|---|---|
| Critical (must fix) | Ghost citations (Error) | — | Likely Fabricated, Cannot Verify | Unsupported | — |
| Warning (should fix) | Orphan references (Warning), name inconsistencies | Error, Warning | Uncertain | Questionable | Error (missing citation) |
| Style (optional) | — | Style | — | — | Warning (citation desert, over-reliance) |
| Agent 3 status | Symbol | Agent 4 status | Symbol |
|---|---|---|---|
| Verified (Zotero/Local/Web) | ✓ | Grounded | ✓ |
| Likely Valid | ✓ | Plausible | ~ |
| Uncertain | ? | Questionable | ? |
| Likely Fabricated | ✗ | Unsupported | ✗ |
| Cannot Verify | — | Unverifiable | — |
Overall column: ✓ Pass if all columns are ✓ or ~. ⚠ Check if any column is ? or —. ✗ Fail if any column is ✗.
Agent 4 works at the claim level, but the Reference-by-Reference table needs a per-reference status. Rollup rule: use the worst claim assessment for that reference. If a reference grounds 5 claims as "Grounded" but 1 as "Questionable", the reference gets "?" in the Grounding column.
health % = (references with ✓ Pass) / (total references) × 100, rounded to nearest integer.
If Agent 2 flags a DOI format issue and Agent 3 flags the same DOI as structurally invalid, keep both findings (they're different checks) but group them on the same row in the Reference-by-Reference table. Do not produce two separate report entries for the same DOI.
Agent 2 provides a corrected version in its "Fix" column. This is for the checker report only — it shows the user what the correct format looks like. The checker does NOT apply fixes. Fixes are applied by /scholar:citation-fixer.
The final output is a markdown report with this structure:
# Citation Verification Report
## Summary
- **Document:** [title/filename]
- **Total in-text citations:** X
- **Total references:** Y
- **Issues found:** Z (N critical, M warnings, P style notes)
- **Overall health:** ██████████░░ 82%
## Critical Issues (must fix)
### Fabricated/Unverifiable References (from Agent 3)
| # | Reference | Status | Details |
|---|-----------|--------|---------|
### Ghost Citations (from Agent 1 — cited but not in reference list)
| # | Citation | Location | Suggested Fix |
|---|----------|----------|---------------|
### Unsupported Claims (from Agent 4)
| # | Claim | Cited Source | Issue | Action |
|---|-------|-------------|-------|--------|
## Warnings (should fix)
### APA 7 Formatting Errors (from Agent 2)
...
### Orphan References (from Agent 1 — in reference list but never cited)
...
### Questionable Attributions (from Agent 4)
...
## Style Notes (optional fixes)
### Missing Citations (from Agent 5)
...
### Over-reliance on Single Sources (from Agent 5)
...
## Reference-by-Reference Status
| # | Reference | Cross-Ref (A1) | Format (A2) | DOI (A3) | Grounding (A4) | Overall |
|---|-----------|----------------|-------------|----------|----------------|---------|
| 1 | Author (Year) | ✓ | ✓ | ✓ | ✓ | ✓ Pass |
| 2 | Author (Year) | ✓ | ✗ | — | ? | ⚠ Check |
...
After presenting the report, offer to run the companion skill /scholar:citation-fixer to apply corrections automatically. The fixer will:
[CITATION NEEDED] markers — at uncited claims, with suggested sourcesThe user can also choose to fix issues manually using the checker report as a guide.
../../references/apa7-guide.md — APA 7 formatting rules (symlink to academic-essay skill)references/verification-patterns.md — Common citation error patterns and how to detect them