Turn your thesis, slides, and committee list into a committee-by-committee defense manual — with an honest weakness audit, reverse-engineered questions, bounded answers, an interactive mock defense, and a readiness score.
Works for any discipline — sciences, engineering, social science, humanities, law, the arts — by detecting your research paradigm and adapting its standards.
English | 中文
🧩 The Problem · 🎯 The Solution · 🧭 How It Works · 🎓 Disciplines · 📄 Output · 📥 Required Inputs · 🚀 Quick Start · 🗂️ Project Layout · 🔌 Compatibility · 🙏 Acknowledgments
The hard part of a defense isn't re-reading your thesis. It's knowing who will ask what, why, and how far your answer can safely go — and finding the weak points before the committee does.
| Common approach | Why it falls short |
|---|---|
| Generic question lists | Too broad; not tied to your committee or your thesis's real weak points |
| Re-reading your thesis | Helps recall, not pressure-handling |
| Asking AI "what might they ask?" | Generic questions, no evaluator context, and the same voice writes the question and a reassuring answer — so it quietly softens the attacks that matter |
| Hiding weak points | You get cornered the moment a committee member presses on evidence |
This skill runs your materials through an adversarial pipeline and ships rehearsal-ready output:
- Detect the research paradigm and judge everything by that field's standards.
- Audit weaknesses first — independently of the committee. A dedicated, attack-only pass produces a read-only Weakness Ledger (severity-ranked, location-cited).
- Reverse-engineer the questions a committee derives from those weaknesses, ranked into a Top-10 most-dangerous list.
- Coach bounded answers that never overclaim (4-move skeleton, 10/30/60-second layers, concede-and-redirect for the fatal ones).
- Research the committee (evidence-graded) to re-weight and personalize — never to fabricate.
- Mock defense: an interactive examiner that keeps attack intensity and doesn't fold when an answer is weak.
- Score readiness (0–100) and ship risk-first output led by a day-of one-pager.
- Generator/evaluator split. The attack and the answer are separate passes. The weakness audit is committed before any answer exists and is read-only afterward — so the questions that can actually sink you don't get quietly smoothed over.
- Weakness-first backbone. Questions come from the thesis's weaknesses first; committee research only re-weights them. Even with zero info on an examiner, the dangerous questions still surface.
- Discipline-adaptive. A paradigm layer swaps the lens: econometric identification for an economics thesis, source criticism for a history thesis, proof validity for a maths thesis — same engine, different lens.
- Anti-overclaiming as
IRON RULES. Simulation isn't sold as reality, correlation isn't causation, future work isn't a finished contribution — and the manual concedes real limits instead of inflating them.
Each stage is backed by a dedicated reference file (the "how"), while SKILL.md stays lean (the "what" + IRON RULES). Re-entry after a revision re-runs only the weakness audit + mock defense.
Detection classifies the paradigm, not the major — a handful of families cover every field.
Six paradigm families — click to expand
| Family | Example fields | Judged by |
|---|---|---|
| Empirical–Quantitative | sciences, engineering, quant social science, finance | design, identification, baselines, reproducibility |
| Empirical–Qualitative | anthropology, sociology, education | positionality, saturation, triangulation, transferability |
| Theoretical–Formal | maths, theoretical CS, analytic philosophy | proof validity, assumptions, non-triviality |
| Textual–Interpretive | literature, history, area studies | source criticism, contextualization, historiography |
| Doctrinal–Normative | law, jurisprudence, normative ethics | doctrinal accuracy, precedent, counterarguments |
| Design–Creative | art, design, architecture, creative writing | craft, concept–artifact link, contribution to practice |
Mixed/interdisciplinary work loads two lenses and watches the seam.
Shipped by risk, not by completeness (default first deliverable = the one-pager):
- Tier A — Day-of one-pager: one-line positioning, readiness score + top-3 exposures, Top-10 questions with 30-second answers, must-say concession lines.
- Tier B — Per-evaluator battle cards: evidence-graded profile + signature questions + the one trap to avoid.
- Tier C — Weakness radar & calibration: severity-ranked ledger + claim-calibration table.
- Tier D — Mock-defense log: drill transcripts, fumbles, re-prioritization.
- Tier E — Full Word manual: the complete by-evaluator guide (
.docx).
A 0–100 readiness score with an exposure tier tells you where to spend limited prep time.
- Your thesis or defense materials (PDF / DOCX / PPTX / Markdown / text / a folder).
- The committee/evaluator list (names required).
- The institution context (school/program; and the defense stage if known).
Ask your agent:
Use thesis-defense-guide to build a defense preparation manual.
My thesis is attached, my committee members are: [names + titles],
and it's a final master's defense at [school / program].
Start with the day-of one-pager and the Top-10 dangerous questions.
Codex: copy the folder into your skills directory:
git clone https://github.com/w1163222589-coder/thesis-defense-guide.git
cp -r thesis-defense-guide ~/.codex/skills/Claude Code / Cowork: place the folder under your skills directory (e.g. ~/.claude/skills/) or install as a plugin, then restart. Any agent that can read files, browse public sources, and run Python can use it.
The bundled
scripts/markdown_to_docx.pyproduces the styled.docx; Markdown output works without it.
Repository layout — 10 reference files + lean orchestrator
.
├── SKILL.md # lean orchestrator: inputs, IRON RULES, Stage 0–7, output tiers
├── references/
│ ├── discipline-profiles.md # paradigm-adaptive lens (6 families) ← read first
│ ├── weakness-audit-framework.md # Three Lenses + Devil's-Advocate dims + severity → Weakness Ledger
│ ├── ppt-audit-checklist.md # slide-level audit (optional, if slides)
│ ├── evaluator-research-protocol.md # evidence-graded research + panel verification
│ ├── question-generation-rules.md # weakness→question engine, escalation, Top-10
│ ├── answer-coaching-framework.md # 4-move bounded answers, stance by severity, anti-overclaim
│ ├── mock-defense-protocol.md # interactive attack-intensity-preserving examiner
│ ├── readiness-rubric.md # 0–100 readiness score + tiers
│ ├── manual-structure.md # Tier-E manual skeleton + highlight labels
│ └── style-rubric.md # tone + Word layout
├── scripts/
│ └── markdown_to_docx.py
├── agents/openai.yaml
├── CHANGELOG.md
├── README.md / README_ZH.md
└── LICENSE
| Platform | Status |
|---|---|
| OpenAI Codex | Supported (local file tools, web research, bundled DOCX converter) |
| Claude Code / Cowork | Supported (equivalent file, research, and Python tooling) |
| Other agents | Adaptable — needs file read, public-source browsing, and Python |
The adversarial-review design draws on concepts from Academic Research Skills by Cheng-I Wu (the Devil's-Advocate review pattern, the "three lenses" review-thinking framework, attack-intensity-preservation, and the cognitive-framework / IRON-RULE style). Concepts were re-implemented and adapted for defense preparation; no text was copied.
Also: Agent Skills, and the reality of thesis-defense pressure — the reason this skill exists.
MIT