Skill: Academic Thesis Review (Chinese Master's Thesis)

Description

Multi-round academic review of Chinese Master's theses (硕士学位论文). Uses a 3-round review strategy (macro → per-chapter → inter-chapter consistency). Acts as a strict professor reviewer, producing actionable revision comments in Chinese. Supports iterative version cycles.

Trigger Phrases

• - "审阅论文" / "review thesis"
• "论文评审" / "学术评审"
• "帮我看看论文" / "论文有什么问题"
• "再做一轮审阅" / "修改后再看看"

Prerequisites

• - Thesis file in .docx format
• Python available for text extraction (zipfile + xml.etree.ElementTree)
• If review_results.md exists in the working directory, iterative mode is triggered automatically (see Phase 1 § Prior Review Detection)

---

Part I: Rules & Instructions

Core Rules

Execution Order

CODEBLOCK0

• - Round 1 and Round 2 have no hard dependency on each other. The agent may run them in parallel or sequentially depending on its own capabilities and context constraints.

- Round 1: reviews the full thesis text. Output: macro structure review + chapter dependency map. - Round 2: reviews chapters sequentially in document order. Each chapter review has access to the summary cards of all previously reviewed chapters. Output: per-chapter issues + per-chapter summary cards.

• - Round 3 depends on both Round 1 and Round 2 completing. It reconciles their outputs, then reviews chapter pairs. Output: inter-chapter consistency issues + pair coverage log.
• All 3 rounds review the same original text — the user revises only after all rounds complete.

Default Thesis Assumption

Standard master's thesis structure: abstract → introduction → literature/theory → methodology → analysis/results → conclusion → references → appendices. No restriction on research topic, industry, or case type.

Non-standard fallback: If the thesis uses non-standard chapter names or merged chapters, map chapters by function: research problem → literature support → method design → results delivery → conclusion.

Conflict Resolution (authoritative — all rounds defer to this)

1. 1. Evidence first, then severity. If R2/R3 evidence overturns an R1 finding, the severity follows the new evidence — do not preserve R1's higher severity for a claim R2/R3 disproved.
2. If Round 2 and Round 3 disagree on the same finding, prefer the conclusion backed by more direct full-text evidence.
3. If multiple rounds independently confirm the same issue but at different severities, keep the higher severity.

Verification Types

All review checklists use these labels. Defined once here; referenced throughout.

• - [D] = directly verifiable from thesis text
• [S] = suspicion only — flag for closer look, agent cannot confirm
• [M] = requires manual verification outside thesis text (original Word file, source literature, or external facts)

Severity Rubric

• - 🔴 Immediate fix — threatens credibility, conclusion validity, or defense safety (data contradictions, unanswered research questions, method-result mismatch, clear factual errors, substantive cross-chapter conflicts)
• 🟡 Strongly recommended — weakens academic quality but not immediately fatal (underdeveloped argument, weak literature-method link, unstable terminology, loose structure)
• 🟢 Quality improvement — wording, polish, or refinement that does not affect conclusions

---

Phase 1: Text Extraction

Extract plain text from .docx via Python:

CODEBLOCK1

Output to review_artifacts/thesis_full_text_<YYYYMMDD_HHMMSS>.txt. Create directory if it does not exist. Note paragraph count and character count as baseline metrics.

Preparation

After extraction, prepare:

• - Full text file — for Round 1
• Per-chapter split files — for Round 2
• Thesis summary block — title + full Chinese abstract + full English abstract + table of contents (~500-1000 chars), used as context prefix in Rounds 2 and 3

Chapter Splitting Rules

• - Split by top-level chapter headings (e.g., 第1章, 第一章, 1 绪论)
• Sub-headings (level 2/3) belong to their parent chapter
• Long chapter: read in subsection chunks during Round 2, but output one consolidated result and one summary card per top-level chapter
• Short chapter: flag as structure weakness in Round 2; do not auto-merge unless user confirms
• Ambiguous boundaries: ask user to confirm rather than guessing

Prior Review Detection (Iterative Mode)

After extraction, check whether review_results.md exists in the current working directory.

Not found → First-version mode:

• - Version = 第1版
• No prior issues to load
• Proceed normally

Found → Iterative mode:

1. 1. Parse the prior review_results.md to extract:

- Prior version number → increment by 1 for the new review
- All prior issues from "本版问题清单" with their severity, location, and description

1. 2. Classify each prior issue by its "位置" field:

- Issues with location "全局" → redistribute to relevant chapters in Round 2
- Issues with location "第X章" → Round 2 (injected in Phase B when reviewing that chapter)
- Issues with location "第X章→第Y章" → Round 3 (injected when reviewing that pair)

1. 3. Verification marking (applied in Phase B of Round 2 and in Round 3):

- Issue is resolved → mark ✅已修改
- Issue is partially addressed → mark ⚠️部分修改, describe what remains
- Issue is unchanged → mark ❌未修改

Iterative Mode: Issue Injection Strategy

Key Principle: Separation of Discovery and Verification

To prevent anchoring bias (agent treating prior issues as a checklist instead of performing independent review), the injection strategy is designed as follows:

Round 1: No injection

• - Execute exactly as first-version mode (completely independent)
• Do NOT inject any prior issues, including "全局" issues
• Rationale: Issues found in Round 1 are typically also discoverable in Round 2/3; verifying global issues through specific chapters is more precise

Round 2: Two-phase per-chapter review

• - Phase A (Independent Discovery): Review chapter with NO historical issues visible
• Phase B (Historical Verification): Inject chapter-specific issues + relevant "全局" issues for verification
• See §2.x for detailed Phase A/B execution rules

Round 3: Cross-chapter verification

• - Inject cross-chapter historical issues (位置="第X章→第Y章")
• Execute spot checks on chapters that declared "无新发现" in Phase A

Anti-Shortcut Rules:

• - ✅ DO: Complete Phase A output before seeing any historical issues. Use dimension coverage checklist to confirm systematic review.
• ❌ DON'T: Combine Phase A and Phase B. Skip the dimension coverage checklist. Claim "no new issues" without checklist evidence.

Extraction Limitations

• - Figures/images lost — only captions and in-text references survive
• Table structure flattened to rows of text
• Formulas may be fragmented or missing
• Cross-references (图X.X, 表X.X) survive as text but actual targets cannot be verified
• Headers/footers and page numbers not extracted

---

Phase 2: Three-Round Review

Round 1: Macro Structure Review

Input: Full thesis text. (In iterative mode: NO prior issues injected — execute as first-version mode)
Focus: Bird's-eye structural and logical assessment. Do NOT give line-level edits.
Long-input fallback: If the full thesis exceeds stable single-pass context, read in large sequential chunks with overlap, then do one pass over all Round 1 findings to ensure cross-chunk consistency before outputting the dependency map.

1.1 Reading Order

1. 1. Title page & metadata
2. Table of contents
3. Chinese abstract + keywords
4. English abstract + keywords
5. All chapters — read sequentially for structural flow
6. References
7. Appendices

1.2 Review Checklist

Item	What to Check	Type
Overall structure	Chapter layout logical? Chapters proportional?	[D]
Research question → Conclusion chain

All research questions answered in conclusion? Any silently dropped? | [D] | | Logic flow | Each chapter leads naturally to the next? Logical jumps or missing bridges? | [D] | | Terminology consistency (global) | Key terms used consistently throughout? | [D] | | Theory utilization | How many theories introduced vs. actually used? Flag bloat. | [D] | | Completeness | All promised analyses delivered? Limitations discussed? Future work? | [D] | | Factual self-consistency | Dates, statistics contradict each other within the thesis? | [D] | | Factual accuracy (external) | Do named facts match known reality? | [M] — flag suspicious values only | | Section proportionality | Chapters too short for their function, or too long and overloaded? | [D] | | Innovation & contribution | Are innovation points explicitly stated? Are they supported by the thesis body? Calibration: master's theses (especially part-time/professional) are NOT expected to produce high-level innovation. Applying existing methods to a new context or combining known approaches in a novel way IS acceptable — the key requirement is that innovation points exist, are clearly articulated, and are not fabricated. Flag missing or unsubstantiated claims, not low novelty. | [D] for whether stated; [S] for whether substantiated | | Practice relevance (if practice-oriented) | Is the research grounded in a real-world problem or case? Does the thesis produce actionable recommendations, not just theoretical conclusions? Does it demonstrate problem-solving ability? | [D] | | Abstract consistency | Chinese and English abstracts match each other and faithfully represent the thesis? | [D] | | References (overview) | Format consistency, recency, relevance | [D] |

1.3 Required Output: Chapter Dependency Map

Output a chapter dependency map at the end of Round 1. This drives Round 3 pair selection.

Format:
CODEBLOCK2

• - A → B = A makes a promise / sets scope; B should deliver or remain consistent
• 🔴 = must check in Round 3 · 🟡 = check if evidence supports it · 🟢 = advisory, low priority
• Principle: 谁承诺，谁兑现，这两章就要配对。 See Appendix A for common patterns.

---

Round 2: Per-Chapter Deep Review

Input per chapter: Thesis summary block + current chapter full text + summary cards of all previously reviewed chapters.
Execution: Sequential, in document order. (In iterative mode: two-phase execution per chapter — see §2.x)
Focus: Detailed intra-chapter quality. Line-level issues are caught here.

First chapter note: When reviewing the first chapter, no prior summary cards exist. Skip the "交付" field in that chapter's summary card.

2.1 Review Dimensions

Dimension	What to Check	Type
Data Consistency	Numbers in tables match text? Actively recalculate totals, percentages, and derived values from raw data — do not just check whether they "look right." If raw numbers are available, verify that reported percentages/scores are arithmetically correct.	[D]
Internal Logic

Argument flows logically? Claims supported? Logical leaps? | [D] | | Methodology Rigor | Method selection rationale stated and internally consistent? | [D]/[S] | | Quantitative checks (if applicable) | Statistical test choices, sample size, validity/reliability — apply only for quantitative empirical work | [S]; definitive judgment [M] | | Literature Review | Research gap articulated? Key competing works cited? Gap justifies this study? | [D] for presence; [M] for citation accuracy | | Citation adequacy | Major claims backed by citations? Unsupported assertions? | [S]; whether source supports claim is [M] | | Language & Style | Typos, repeated chars (的的/了了), grammar errors (病句), punctuation misuse (中英文标点混用), inconsistent terminology, colloquial language in academic text, overly long/convoluted sentences | [D] | | Factual accuracy (internal) | Dates/numbers consistent within this chapter? | [D] | | Factual accuracy (external) | Real-world claims plausible? | [S]/[M] — flag only | | Chapter Framing | Opening bridges from previous chapter? Closing sets up next? | [D] | | Chapter Proportion & Function | Too short for its function? Too long and overloaded? | [D] |

Proactive expansion rule: When the thesis uses a specific research method (survey, AHP, regression, case study, grounded theory, etc.), the agent must proactively apply that method's academic validation criteria — not just check the generic dimensions above. The agent's own domain knowledge is the source for method-specific checks; the skill does not enumerate them.

See Appendix B for chapter-type-specific focus areas.

2.2 Special Section Responsibilities

• - Abstracts — Round 2: wording quality, terminology stability, grammar; whether abstract claims overstate the body
• References — Round 2: flag uncited claims or mismatched citations [S]; whether cited source supports the claim is [M]
• Appendices — Treat as a distinct Round 2 unit: questionnaires, interview guides, supplementary tables, coding schemes

2.x Iterative Mode: Two-Phase Per-Chapter Review

Applies only when review_results.md exists (iterative mode).

In iterative mode, each chapter review is split into two sequential phases to prevent anchoring bias:

Phase A: Independent Discovery (NO historical issues visible)

Input: Thesis summary block + current chapter full text + prior chapter summary cards

Output format (to review_working.md):
CODEBLOCK3

Phase B: Historical Issue Verification (inject after Phase A completes)

Input: Phase A output + historical issues for this chapter (including relevant "全局" issues)

Output format (to review_working.md):
CODEBLOCK4

Sequencing Rule: Phase A MUST be complete and output to working file BEFORE Phase B input is provided. The agent may NOT "peek ahead" at historical issues during Phase A.

Dimension Coverage Checklist Requirement: The checklist in Phase A must include specific descriptions of what was checked (not just checkmarks). This serves as verifiable evidence that independent review was performed.

2.3 Required Output: Per-Chapter Summary Card

Output after each chapter. Used by Round 3 for orientation and dependency map validation.

CODEBLOCK5

During Round 3 setup, compare summary cards against Round 1's dependency map:

• - Summary shows dependency Round 1 missed → add to Round 3
• Round 1 mapped dependency no summary confirms → demote or remove

Conflict with Round 1: default to Round 2 evidence (per Core Rules § Conflict Resolution).

---

Round 3: Inter-Chapter Consistency Review

Input per pair: Thesis summary block + both summary cards + both chapter full texts. In iterative mode: also receives cross-chapter prior issues for verification.
Pairs: Merged from Round 1 dependency map + Round 2 summary card validation.

Pair Selection Stopping Rule

• - 🔴 pairs — mandatory
• 🟡 pairs — review only if summary cards show concrete promise/deliverable link or data/terminology risk; skip otherwise
• 🟢 pairs — skip unless user requests exhaustive mode

Coverage Transparency

Output at end of Round 3:
CODEBLOCK6

Spot Check Mechanism (Iterative Mode Only)

Purpose: Validate Round 2 Phase A "无新发现" claims through targeted re-examination. Provides a deterrent against superficial review.

Trigger Conditions:
A spot check is triggered when ALL of the following are true:

1. 1. A chapter's Phase A declared "本章未发现新问题"
2. AND that chapter has historical 🔴 issues OR 3+ historical 🟡 issues

Procedure:

1. 1. Select 1-2 subsections (500-1000 characters each) from the triggered chapter
2. Perform focused re-review on two dimensions only: data consistency + terminology consistency
3. Output result to INLINECODE13

Limits (Cost Control):

• - Maximum 3 spot checks per review cycle
• Each spot check covers 500-1000 characters only
• Focus on 2 dimensions only (data + terminology)

Output Format (to review_working.md):
CODEBLOCK7

Integration with Final Output:

• - Spot check results are recorded in review_working.md only
• If spot check finds omissions, add them to the consolidated issue list with tag INLINECODE16
• INLINECODE17 format remains unchanged (user does not see spot check details)

Reading Strategy (per pair)

Full-text reading of both chapters is MANDATORY. Summary cards are navigation index only.

1. 1. Read both summary cards → understand expected promise/deliverable relationship and key data points
2. Read Chapter A in full → note claims, numbers, definitions, commitments
3. Read Chapter B in full → actively cross-check against Chapter A notes
4. Document all inconsistencies found

Long-input fallback: If a chapter pair exceeds stable single-pass context, read each chapter in chunks but ensure a final cross-check pass covers all noted data points and commitments from both chapters before concluding.

Pre-Round 3: Dependency Map Reconciliation

1. 1. List Round 1's dependency pairs
2. Check each summary card's "承诺"/"交付" against Round 1's pairs
3. Produce final pair list: confirmed + newly discovered − demoted
4. Document changes briefly

Review Checklist (per pair)

Item	What to Check	Type
Terminology alignment	Same concepts use same terms across chapters?	[D]
Data/number consistency

Same statistics, sample sizes, dates consistent? | [D] | | Scope alignment | Chapter B delivers what A promised? B introduces undeclared scope? | [D] | | Cross-reference suspicion scan | Suspicious 图/表/章/节 references or numbering jumps? | [S]/[M] | | Logical continuity | Transition makes sense? Logical gaps or contradictions? | [D] |

Cross-references are not a primary hard-check item — only flag suspicious cases for manual Word verification.

---

Phase 3: Consolidation & Output

After all 3 rounds, consolidate into a single issue-centric document. Each issue appears once; originating round noted in brackets.

Output Files

File	Location	Content
INLINECODE18	Working directory	Final consolidated review (user-facing deliverable)
INLINECODE19

review_artifacts/ | Round 1 dependency map, Round 2 summary cards, Round 3 pair logs, all intermediate findings | | thesis_full_text_<YYYYMMDD_HHMMSS>.txt | review_artifacts/ | Extracted plain text from docx (Phase 1 output) |

Directory structure:
CODEBLOCK8

Notes:

• - Create review_artifacts/ directory if it does not exist
• Timestamp format: YYYYMMDD_HHMMSS (local time)
• Working file serves as execution evidence and traceability record
• In iterative mode, each review cycle creates a new timestamped working file

Working File Structure (Iterative Mode)

In iterative mode, review_working.md must include the following sections to serve as execution evidence:

CODEBLOCK9

Key Points:

• - Phase A and Phase B records must be clearly separated
• Dimension coverage checklist must include specific descriptions (not just checkmarks)
• Spot check records (if any) must be included
• This structure provides traceability for the review process

Consolidation Rules

De-duplication: Keep the most specific/actionable version of each issue. Other rounds noted as supporting evidence only. When merging related-but-distinct issues, verify they truly share the same root problem — two issues that share a topic (e.g., both about Chapter 1) but target different defects (e.g., tone vs. evidence gap) must remain separate entries.

Conflict resolution: See Core Rules § Conflict Resolution.

Coverage verification (mandatory before finalizing): Before writing the final output, confirm that every distinct finding from all 3 rounds is either included in the consolidated issue list, explicitly merged into another issue, or dropped with a stated reason. No finding may be silently lost.

Iterative mode — prior issue handling:

• - All prior issues verified by the rounds are collected here
• Issues marked ✅已修改 go into "修订情况核查" only (not into the problem list)
• Issues marked ⚠️部分修改 go into both "修订情况核查" and the problem list (tagged [部分修改])
• Issues marked ❌未修改 go into both "修订情况核查" and the problem list (tagged [延续])
• Newly discovered issues in this review are tagged [新发现] in the problem list

Persistent issues: If the same issue has been flagged across 3+ consecutive versions:

1. 1. Acknowledge explicitly ("已连续N版提出")
2. Provide minimum-cost fix — exact replacement text + location
3. Assess defense risk
4. Version 4+: shift to "if you cannot fix, prepare an oral answer for defense"

Convergence criteria (assess and report in 总体评价):

• - [ ] Zero 🔴 items
• [ ] All 🟡 either fixed or accepted with defense prep
• [ ] No new issues in full 3-round review
• [ ] Suspicious cross-references flagged for manual Word check

Output Template

Template instructions (do NOT include these in the final document):

• - §一 修订情况核查: include only in iterative mode (prior review_results.md exists). Omit entirely in first-version mode.
• §二 问题清单: contains ALL current open items (unfixed prior + partially fixed prior + newly discovered). In first-version mode, all issues are implicitly [新发现] and the tag may be omitted.
• §三 收敛评估: include only in iterative mode.
• [S] items must use uncertain wording ("疑似…""建议核查…"); [M] items must state what needs manual verification.

CODEBLOCK10

---

Output Specification

• - Language: Chinese, with English terms where academically standard
• Format: Markdown (.md), filename INLINECODE31
• Tone: Strict but constructive — every criticism paired with a specific fix suggestion
• Priority: 🔴🟡🟢 consistently used across all rounds and versions

---

Part II: Reference Knowledge

The following sections are reference material for the agent to consult during review. They are not execution rules.

---

Appendix A: Default Dependency Patterns

Common dependency pairs for standard thesis structure. Use as starting point for Round 1 dependency map.

Pair	What to Check
Introduction → Literature Review	Research questions fully covered by literature? Irrelevant literature present?
Introduction → Conclusion

Every research question answered? Conclusion claims within intro scope? |
| Literature Review → Methodology | Conceptual framework operationalized in method? Method justified by gap? Terminology consistent? |
| Methodology → Results | Every described method/instrument executed and reported? Undescribed methods appearing? Variable names consistent? |
| Results → Discussion/Conclusion | Every finding discussed? Over-interpretation beyond data? Statistical conclusions match data? |

---

Appendix B: Chapter-Specific Focus Areas

Additional focus when reviewing specific chapter types in Round 2.

Chapter Type	Additional Focus
Introduction	Research gap specificity, innovation points clearly stated and not overstated, stated framework matches actual thesis structure
Literature Review / Theory

Theory bloat (describe 5, use 2), citation format, gap analysis → methodology link |
| Methodology | Rationale stated, internal consistency; (if quantitative) sample justification, validity/reliability, reproducibility |
| Results / Analysis | Table-text data consistency; (if quantitative) test choices, hypotheses; (if qualitative) interpretation grounded in evidence |
| Discussion / Conclusion | No over-interpretation, all questions answered, limitations honest; (if practice-oriented) actionable recommendations for practitioners, not just theoretical implications |
| English Abstract | Grammar, non-idiomatic phrasing ("Aiming at", "Firstly/Secondly"), keyword consistency |
| Appendices | Alignment with main text; (if survey) question design, scale coverage, response options |

---