Auditoria de Citações em Teses
ekontoTURBO/thesis-citation-audit
Valida citações APA em teses contra fontes PDF com relatório automático
Ver no GitHubSkill que verifica automaticamente todas as citações APA em uma tese (.docx) contra PDFs de fontes, frase por frase, usando sub-agentes paralelos. Gera relatório detalhado em markdown com resumo executivo, verificações por seção e lista priorizada de correções.
Instalação
npx convez-ai ekontoTURBO/thesis-citation-audit1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Validar integridade de citações em teses de conclusão de curso
- Detectar citações fabricadas, orfãs ou com dados incorretos
- Identificar desajustes entre bibliografia e texto (anos, autores, páginas)
Veredito

Ferramenta especializada e robusta para pesquisadores que precisam validar grandes volumes de citações. Detecta erros sutis (desvios de página, autores incorretos, fontes fabricadas) que ferramentas genéricas perdem. Requer setup técnico (Poppler, python-docx) mas economiza semanas de revisão manual.
README
A tradução para o português deste README ainda está na fila.
thesis-citation-audit
A Claude Code skill that verifies every APA citation in a thesis (.docx) against the source PDFs, sentence-by-sentence, using parallel sub-agents. Produces a master markdown report with executive summary, per-section verifications, and a prioritized fix list.
Built from a workflow that audited 4 successive versions of a real bachelor's thesis (~276 citations across 19 subsections) and caught:
- Fabricated page numbers (citing s. 257-270 of an article that ends on s. 155)
- Wrong authors (e.g. "Hayden 2019" when the author is actually Gabriel Moss)
- Orphaned citations not in the bibliography
- Source files that don't match the declared work (PDF is a review, not the book; preprint vs published pagination)
- Systematic page-number drift (e.g. all Author X citations −1 page)
- Bibliography/text mismatches (wrong year, missing co-author, wrong journal)
Install
Option A — User-level (recommended)
git clone https://github.com/ekontoTURBO/thesis-citation-audit ~/.claude/skills/thesis-citation-audit
# or on Windows PowerShell:
git clone https://github.com/ekontoTURBO/thesis-citation-audit "$env:USERPROFILE\.claude\skills\thesis-citation-audit"
Then in Claude Code, invoke with /thesis-citation-audit or describe the task in natural language ("verify the citations in my thesis").
Option B — Project-level
Copy this folder to ./.claude/skills/thesis-citation-audit/ inside the project, or:
git clone https://github.com/ekontoTURBO/thesis-citation-audit .claude/skills/thesis-citation-audit
A sample anonymized report excerpt lives in examples/sample_report_excerpt.md.
Dependencies
pip install python-docx
# Poppler (provides pdftotext):
# macOS: brew install poppler
# Ubuntu: sudo apt install poppler-utils
# Windows: choco install poppler (or use Git Bash bundled poppler)
Verify:
python -c "import docx; print('python-docx OK')"
pdftotext -v
Usage
In Claude Code:
/thesis-citation-audit "C:\path\to\thesis.docx" "C:\path\to\sources_folder" "C:\path\to\REPORT.md"
Or just say: "verify the citations in my thesis at <path>, sources in <folder>" — Claude will ask for any missing arguments.
The skill:
- Extracts thesis paragraphs and detects structure (chapters, subsections, bibliography, netografia).
- Inventories every APA citation (parenthetical + narrative).
- Splits thesis into per-subsection files.
- Dispatches one sub-agent per subsection in parallel. Each verifies every citation against the source PDFs.
- Consolidates the per-section reports into one master markdown file.
- Writes an executive summary with critical issues, patterns, and prioritized fixes at the top.
Wall-clock time: 15–45 minutes for a typical 60–80 page thesis. Token cost scales with (# subsections × avg citations × avg PDF size).
What the report looks like
# CITATION AUDIT REPORT — thesis_v10
## EXECUTIVE SUMMARY
### 🚨 Problems CRITICAL
1. "Kujawski, 2024" [2.1, P0193] — orphaned. Actual author: Monika Czerwonka (Obserwator Finansowy).
2. "Hopper, D. (2025)" [3.4, P0297] — wrong author. Actual: Henry Stockdale, 6 Sept 2024 (UploadVR).
...
### 🔴 Bibliography mismatches
1. Thaler & Sunstein 2008/Yale in bibliography vs in-text 2021/Penguin Final Edition.
...
# Section 1.1 — Pojęcie sprzedaży
## [P0128] (Bagozzi, 1975, s. 32)
**Status:** ✅ OK
**Comment:** Direct match on printed s. 32.
## [P0129] (Encyklopedia Zarządzania, 2026)
**Status:** ❌ TREŚĆ + ❓ BRAK ŹRÓDŁA
**Comment:** File contains entry "Wymiana", not "Sprzedaż". The cited phrase doesn't appear in the file.
...
Architecture
thesis-citation-audit/
├── SKILL.md # main skill instructions (under 500 lines)
├── README.md # this file
├── scripts/
│ ├── extract_thesis.py # docx → text + structure + citation inventory
│ ├── split_sections.py # per-subsection files
│ └── merge_report.py # consolidate sub-agent reports
└── templates/
├── section_prompt.md # exact prompt template for sub-agents
└── exec_summary.md # executive summary structure
All scripts are stdlib-only except python-docx. No API keys required.
Languages
Sub-agents mirror the thesis language. Polish thesis → Polish report. English thesis → English report. Status legend uses bilingual labels (⚠️ STRONA / PAGE).
What it does NOT do
- It does not check non-APA citation styles (Chicago, MLA, Harvard footnotes). Targeted at author-year APA.
- It does not edit the docx. You get a report; the author decides what to fix manually.
- It does not invent citations. If a PDF can't be opened, the citation is flagged ❓ — never silently guessed.
- It does not check plagiarism (use Turnitin/iThenticate). It checks attribution accuracy.
License
MIT. Use freely, fork freely.
Credits
Built with Claude Code by Eryk Czekalski. Forged in the fire of 4 thesis revisions.