convez.ai
SkillEscrita & ConteúdoIntermediário
ibrahim-kukash

Thesis Writer

ibrahim-kukash/thesis-writer

Escreve teses acadêmicas que passam em detectores de IA e examinadores

Ver no GitHub
BomPluginScore · de 0 a 100

Skill para escrever capítulos de tese que passam em verificações de detecção de IA. Detecta 14 padrões de escrita artificial (baseado em framework de Wikipedia), escaneia textos existentes, corrige marcadores de IA preservando significado, e prepara para viva com +35 perguntas categorizadas.

Instalação

$npx convez-ai ibrahim-kukash/thesis-writer

1 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

Estrelas
16
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Escrever capítulos de mestrado/doutorado com qualidade de publicação
  • Verificar teses existentes contra 14 padrões de escrita artificial
  • Preparar-se para viva com banco de perguntas de examinadores
  • Corrigir textos sinalizados mantendo rigor acadêmico e citações

Veredito

Ferramenta especializada e bem fundamentada para acadêmicos que precisam passar por anti-AI detection. Framework de 14 padrões é sólido (baseado em Wikipedia), mas requer que o usuário entenda nuances de escrita acadêmica para aproveitar as correções. Cobre genuinamente um problema real no ecossistema de teses contemporâneas.

README

A tradução para o português deste README ainda está na fila.

thesis-writer

Write publication-quality MSc/PhD thesis chapters that pass examiner scrutiny and AI detection tools.

What it does

  • Writes thesis chapters following academic conventions, with proper hedging, citations, and chapter structure
  • Scans existing text for 14 AI writing patterns (based on Wikipedia: Signs of AI writing)
  • Fixes flagged patterns while preserving meaning and citations
  • Prepares for viva with 35+ categorised examiner questions

Install

# Claude Code
claude skill add /path/to/thesis-writer

# Or clone and point to it
git clone https://github.com/futureweapons2022-crypto/thesis-writer.git ~/.claude/skills/thesis-writer

Usage

# In Claude Code
/thesis-writer write chapter 2 section 2.1

/thesis-writer scan my-chapter.docx

/thesis-writer prepare for viva

Standalone AI scanner

python scripts/ai_scan.py thesis_draft.docx
python scripts/ai_scan.py --text "paste your text here"

Output:

============================================================
THESIS AI SCAN REPORT
============================================================
Score: 72/100
Words: 2515 | Sentences: 89 | Paragraphs: 29
Violations: 2 critical, 3 medium, 1 minor

VIOLATIONS:

  ❌ [CRITICAL] Pattern 1: Banned vocabulary
    Found: "comprehensive" (2x)
    Fix: Replace with a plain alternative

  ⚠️ [MEDIUM] Pattern 2: However chain
    Detail: 5 transition starters in 2515 words (2.0 per 1000)
    Fix: Use subordinate clauses or plain conjunctions

What's inside

thesis-writer/
├── SKILL.md                        # Main skill (3 modes: Write, Scan, Fix)
├── README.md
├── LICENSE
├── references/
│   ├── ai-patterns.md              # 14 patterns to avoid + 9 human-writing signals
│   ├── chapter-rules.md            # Chapter-by-chapter writing rules
│   └── viva-questions.md           # 35+ examiner questions by difficulty
└── scripts/
    └── ai_scan.py                  # Standalone AI pattern scanner (stdlib only)

The 14 AI patterns it catches

#PatternSeverityExample tell
1Banned vocabularyCritical"delve," "comprehensive," "leverage"
2However chainCriticalHowever, / Moreover, / Furthermore, back-to-back
3Uniform sentence lengthCriticalEvery sentence 20-25 words
4Copula avoidanceMedium"serves as" instead of "is"
5Superficial -ing endingsMedium"highlighting the importance of..."
6Rule of threeMinor"X, Y, and Z" in every paragraph
7Synonym cyclingMediumSame concept, different words each time
8Negative parallelismsMinor"Not just X, but also Y"
9False rangesMinor"From X to Y, from A to B"
10Em dash overuseMinorMore than 2 per page
11Filler phrasesMinor"It is worth noting that..."
12Formulaic endingsMedium"These findings highlight the importance..."
13Generic attributionMedium"Several studies have shown..."
14Descriptive lit reviewCriticalListing studies without comparing them

Anti-AI detection approach

Based on how detectors work:

  • Perplexity: AI text is too predictable. The skill forces less obvious word choices.
  • Burstiness: AI sentences are too uniform. The skill enforces dramatic length variation (5-word and 40-word sentences in the same paragraph).

Non-native English speakers face 61% false positive rates in AI detectors. This skill is especially useful for international students.

Research basis

The patterns are grounded in published work — and deliberately updated as the models change:

  • Lexical tells (the banned-word list) come from Kobak et al., Delving into LLM-assisted writing in biomedical publications through excess vocabulary, Science Advances 2025 (arXiv:2406.07016) — a 14M-abstract study of which words surged after ChatGPT (delves ×25, showcasing ×9, underscores ×9).
  • These tells decay. "delve" peaked in 2023–24 and faded through 2025 as labs trained it out, so current-gen models lean on formulaic transitions instead — "In conclusion," "It is important to note," "Notably" — flagged separately (pattern 29).
  • The durable signals are structural, not lexical: sentence-length burstiness, clause-density variation, and punctuation entropy (pattern 28). Stylometric research (arXiv:2511.21744, arXiv:2510.00890) finds combined structural features reach F1 ≈ 0.94, well above perplexity alone — and unlike word lists, they don't expire.

Honest caveat: even the best stylometric detectors top out at ~80–95% accuracy and drop when text is edited. Treat this as a guide, not proof.

Credits

License

MIT

Ver o original no GitHub

Bom para

Relacionados