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

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

• AI patterns based on Wikipedia: Signs of AI writing by WikiProject AI Cleanup
• Viva questions compiled from University of Calgary, Oxford, and Edinburgh examiner guides
• Skill structure follows the claude-skills authoring standard

License

MIT