agentsclimarketplace

Thesis

Skill VictorAurelius/claude-starter-kit/skills/document-generation/thesis

Use khi user nói 'làm khóa luận', 'thesis V1', 'audit thesis', 'thesis ready to defense?', 'render thesis docx', hoặc trước khi ship academic deliverable. Workflow: planning → draft chapters → 3-agent outside-in audit → fix → assemble DOCX. Backed by thesis-content-standard.md rubric /100.From its SKILL.md

Install
npx -y skills add VictorAurelius/claude-starter-kit --skill thesis

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
  • runs commandsInstructs the agent to run 8 commands, including `mkdir -p documents/<thesis-dir>/{chapters,references,figures,audits}` and 7 more.

SKILL.md

14.7 KB, ~4.2k tokens by cl100k_base, as published. Nobody here has run it

Thesis — Vietnamese-University Bachelor/Engineer/Master Khóa Luận Workflow

Skill này hướng dẫn workflow ship academic deliverable (khóa luận tốt nghiệp) cho dự án phần mềm thực tế, chấm theo rubric .claude/rules/thesis-content-standard.md /100 (9 categories).

Khi nào dùng

  • Bạn dùng dự án thực tế (production code, 200+ commits) làm thesis topic
  • Cần ship thesis-v1.docx cho academic submission (cử nhân / kỹ sư / thạc sĩ)
  • Đại học VN: convention UTC/HUST/UET/HCMUT/BKHN (rule là opinionated cho VN scope)
  • Universities khác: skill workflow vẫn áp dụng, rule §C1 cần adapt page-size + heading numbering

Output

  • 4-6 chapter MDs trong documents/<thesis-dir>/ đạt rubric ≥75/100 C+
  • thesis-v1.docx assembled từ chapter MDs (python-docx pipeline OR pandoc)
  • 3 outside-in audit reports trong documents/<your-audits-dir>/persona-review/
  • Bibliography IEEE format ≥30 entries (cử nhân) / ≥50 (kỹ sư) / ≥80 (thạc sĩ)

Setup for your project

1. Create thesis directory structure

mkdir -p documents/<thesis-dir>/{chapters,references,figures,audits}
cp .claude/skills/document-generation/thesis/templates/thesis-info.md.template \
   documents/<thesis-dir>/thesis-info.md
cp .claude/skills/document-generation/thesis/templates/chapter-mapping.md.template \
   documents/<thesis-dir>/chapter-mapping.md
cp .claude/skills/document-generation/thesis/templates/bibliography.md.template \
   documents/<thesis-dir>/references/bibliography.md

2. Fill thesis-info.md

Update with your details:

  • Student name, MSSV, lớp, khóa
  • GVHD (advisor) name + title
  • GVPB (reviewer) name + title (nếu đã được assign)
  • School + faculty + department
  • Thesis level (cử nhân / kỹ sư / thạc sĩ)
  • Target page count per rule §4 cap
  • Submission deadline + defense window

3. Customize rule per school

Edit .claude/rules/thesis-content-standard.md §C1 if your school's spec differs from VN baseline:

  • Page size (default A4)
  • Margins (default T=2.5 B=2.5 L=3.0 R=2.0 cm)
  • Heading sizes (default TNR 13/14/16/18pt body/H3/H2/H1)
  • Numbering convention (default CHƯƠNG N. / N.M / N.M.P)
  • Bibliography style (default IEEE)

Workflow

Phase 1 — Planning (1-2 sessions)

Goal: chapter outline + bibliography seed + outside-in pre-audit

Step 1.1: Draft chapter outline

Create documents/<thesis-dir>/chapter-mapping.md per template. VN cử nhân CNTT typical:

  1. Chương 1 — Tổng quan (background + scope + literature)
  2. Chương 2 — Cơ sở lý thuyết + Phân tích yêu cầu (theory + business analysis)
  3. Chương 3 — Thiết kế hệ thống (architecture + diagrams)
  4. Chương 4 — Triển khai và Kết quả (implementation + KPI)
  5. KẾT LUẬN VÀ KIẾN NGHỊ (conclusion + future work + đóng góp khoa học)

NOT: ghép Phân tích + Thiết kế vào 1 chương (anti-pattern per rule §C2).

Step 1.2: Outside-in pre-audit (recommended)

Run 3-agent simulation BEFORE writing chapters:

# Spawn 3 parallel agents simulating thesis committee
# Agent 1: GVHD (advisor) — focuses on academic content + methodology
# Agent 2: GVPB (reviewer) — focuses on legal/compliance + presentation
# Agent 3: Defense committee — focuses on examiner-question failure modes

Use prompts from reference/outside-in-audit-prompts.md. Findings inform chapter structure BEFORE 1000+ lines written.

Step 1.3: Bibliography seed

Pre-populate references/bibliography.md với:

  • Your department's giáo trình (1-2 refs minimum — per rule §C3)
  • VN law refs nếu scope touches data/privacy (PDPL 2023, Decree 13/53/2022, Thông tư 78/2021)
  • Peer-reviewed papers (≥30% per rule §C3)
  • Vendor docs OK but weight thấp

Phase 2 — Draft (3-5 sessions)

Goal: chapter content MDs đạt baseline rubric ≥60/100 (C grade ok at this stage)

Step 2.1: Write chapters in Vietnamese

PHẢI follow rule §S3 (Vietnamese narrative strict):

  • Headings 100% tiếng Việt (NOT "Domain capabilities" → "Năng lực miền")
  • Body prose-style (NOT listing pattern ≥5 items single sentence)
  • English technical token natural (JWT, OIDC, SaaS) với parenthetical define lần đầu

Step 2.2: Cite as you write

Per rule §S4 — every numeric/factual claim needs [N, tr.NNN]:

  • Direct quote "..." → page-num mandatory
  • Specific stat ("80% trung tâm chuyển khoản") → page-num mandatory
  • Vendor scale ("MISA EMIS 30,000 trường") → cite vendor URL + access date

Step 2.3: Measurement methodology block

Per rule §S5 — every performance claim needs:

  • Tool (Lighthouse / JMeter / wc -l)
  • N (sample size)
  • Date (YYYY-MM-DD HH:MM TZ)
  • Env (endpoint / network condition / hardware)

Step 2.4: Define acronyms at first use

Per rule §S7 — GVCN (giáo viên chủ nhiệm), OIDC (OpenID Connect, giao thức xác thực mở), DPO (Data Protection Officer, cán bộ bảo vệ dữ liệu).

Phase 3 — Outside-in audit (1-2 sessions)

Goal: surface 30-80 findings inside-out review misses

Step 3.1: Spawn 3 parallel audit agents

Agent 1 (Persona Simulation):
  Role-play GVHD reading thesis V1 cover-to-cover
  Output: 10-15 findings (academic depth, methodology rigor, contribution clarity)

Agent 2 (UTC Benchmark / Your-school Benchmark):
  Compare against known-good DOCX sample from your school's archive
  Output: 10-15 findings (format compliance, convention match)

Agent 3 (Failure-mode Matrix):
  Defense committee chất vấn simulation — top 10 fail modes
  Output: 10-15 findings (examiner-question vulnerabilities)

Combine → typical 30-50 findings beyond inside-out review.

Step 3.2: Apply rubric scoring

Score each chapter MD per rule §2 9-category rubric:

CategorySelf-scoreTargetGap
C1 Format/15≥13 PASS
C2 Content + page count/15≥13 PASS
C3 Bibliography/15≥13 PASS
C4 Academic tone/15≥12 PASS
C5 Project-internal scrub/10≥8 PASS
C6 Draft-marker scrub/5≥4 PASS
C7 Diagram + figure/10≥8 PASS
C8 Examiner readiness/10≥8 PASS
C9 Compliance + legal/5≥4 PASS
Total/100≥75 C+

Step 3.3: Prioritize fixes

  • Hard-FAIL items first (C5 "Claude" mentions / C9 "vi phạm" admissions — auto category FAIL)
  • Page count over cap (C2 — bachelor >90 trang auto-FAIL)
  • Diagram rendering (C7 — Mermaid → PNG pipeline if not yet wired)
  • Then incremental polish (C4 tone + C6 draft markers)

Phase 4 — Fix sweep (2-3 sessions)

Goal: apply audit findings retroactively to chapter MDs

Use rule §3 banned-patterns table as grep targets:

# C5 — project-internal references
grep -rn "Claude\|Wave [0-9]\|Phase [0-9]\|GAP-[0-9]\|\.claude/" \
  documents/<thesis-dir>/chapters/*.md

# C6 — draft markers
grep -rn "## TL;DR\|TODO\|FIXME\|\[placeholder\]\|\[stub\]" \
  documents/<thesis-dir>/chapters/*.md

# C4 — emoji + slang
grep -rnP "[\x{1F300}-\x{1F9FF}]|✅|❌|⚠️|🎉|🚀|📅|🆘" \
  documents/<thesis-dir>/chapters/*.md

# C4 — banned word "đối thủ"
grep -rn "đối thủ" documents/<thesis-dir>/chapters/*.md

Each pattern → replace per rule §3 table.

Phase 5 — Assemble DOCX (1 session)

Goal: chapter MDs → final thesis-v1.docx

Option A — Python-docx pipeline (recommended for fine control)

Use template at templates/thesis-assemble.py.template:

python3 .claude/skills/document-generation/thesis/templates/thesis-assemble.py.template \
  --chapters documents/<thesis-dir>/chapters/ \
  --bibliography documents/<thesis-dir>/references/bibliography.md \
  --info documents/<thesis-dir>/thesis-info.md \
  --output documents/<thesis-dir>/thesis-v1.docx

Customize template for:

  • School logo path
  • Cover page layout (bìa cứng + bìa phụ)
  • Heading style mapping (H1→Heading 1 TNR 18pt, H2→Heading 2 TNR 16pt, etc.)
  • Mermaid → PNG conversion via headless browser (mermaid-cli)
  • Bibliography IEEE rendering với hanging indent

Option B — Pandoc pipeline (faster setup, less control)

pandoc \
  --from gfm \
  --to docx \
  --reference-doc=.claude/skills/document-generation/thesis/templates/thesis-reference.docx.template \
  --metadata-file documents/<thesis-dir>/thesis-info.md \
  --filter pandoc-mermaid \
  --citeproc \
  --bibliography documents/<thesis-dir>/references/bibliography.bib \
  --csl ieee.csl \
  -o documents/<thesis-dir>/thesis-v1.docx \
  documents/<thesis-dir>/chapters/*.md

Requires: pandoc, pandoc-mermaid (Mermaid filter), CSL file (IEEE style từ https://www.zotero.org/styles).

Pre-ship verification

# Verify A4 page size + margins
python3 -c "
from docx import Document
d = Document('documents/<thesis-dir>/thesis-v1.docx')
s = d.sections[0]
assert s.page_width.cm == 21.0, f'Page width {s.page_width.cm} != 21.0'
assert s.page_height.cm == 29.7, f'Page height {s.page_height.cm} != 29.7'
print('A4 page size OK')
"

# Verify zero Mermaid code blocks leaked (means rendered as PNG)
unzip -p documents/<thesis-dir>/thesis-v1.docx word/document.xml | grep -c '```mermaid' && echo 'FAIL: Mermaid leaked' || echo 'Mermaid → PNG OK'

# Page count
libreoffice --headless --convert-to pdf documents/<thesis-dir>/thesis-v1.docx --outdir /tmp/
pdfinfo /tmp/thesis-v1.pdf | grep Pages

Phase 6 — Defense prep (1-2 sessions)

Goal: committee Q&A drill

Step 6.1: Anticipate top 10 questions

From outside-in failure-mode audit (Phase 3 Agent 3). Common categories:

  • Methodology rigor ("Vì sao chọn approach X thay vì Y?")
  • Compliance ("PDPL DPO khi nào engage?")
  • Empirical evidence ("KPI 92/100 Lighthouse — measure thế nào?")
  • Scope ("Vì sao không include Z?")
  • Contribution ("Đóng góp khoa học cụ thể là gì?")

Step 6.2: Slide deck (10-15 slides max)

Structure:

  1. Title + student info
  2. Bối cảnh + motivation (1-2 slides)
  3. Phương pháp luận (1 slide)
  4. Kiến trúc + design highlights (3-4 slides)
  5. Kết quả + KPI (2-3 slides)
  6. Đóng góp khoa học (1 slide — explicit per rule §C2)
  7. Hạn chế + Hướng phát triển (1 slide)
  8. Q&A + cảm ơn

Step 6.3: Print kit

  • Thesis hard-cover binding (gáy in "HỌ TÊN - LỚP - NĂM" per rule §C1)
  • Slide handout 3-up
  • Backup USB với DOCX + PDF + supporting figures

Gotchas (from real shipping)

Page count creep

Bachelor cap = 90 trang auto-FAIL. Real ship often 110+ trang draft → forced trim. Mitigation: target 70 trang draft, leave 20-trang headroom for polish bloat.

Mermaid rendering

GitHub native render Mermaid in markdown, BUT python-docx pipeline strips code blocks to text by default. Mitigation: pre-render Mermaid → PNG via mmdc (mermaid-cli) headless browser BEFORE chapter MD ingestion. Inject ![Hình X.Y](mermaid-output/figure-N.png) image refs.

Bibliography drift

Body cites [15] but bibliography numbering shifts when you add new ref → broken cite. Mitigation: use BibTeX + CSL with pandoc (Option B above) for auto-numbering. OR maintain bibliography append-only (new refs ALWAYS appended at end, never reorder).

Project-internal language leakage

You say "Claude" / "Wave 5" / "GAP-123" naturally in conversation — they leak into thesis body. Mitigation: grep §3 banned patterns EVERY session before commit. Rule §7.4 detector script catches.

Compliance phrasing

Easy trap: "Phase 1 BETA vi phạm Decree 53 nhưng acceptable" — committee will catch. Mitigation: rewrite mềm "Phase BETA invite-only chưa kích hoạt ngưỡng Decree 53 §26 (1M user threshold); roadmap migrate sang AWS Hanoi Local Zone OR VN cloud trong giai đoạn GA".

Diagram source attribution

VN spec §2.4 mandates *Nguồn:* attribution per figure. Easy to miss for self-built diagrams. Mitigation: auto-inject *Nguồn: tác giả tự xây dựng* for any Mermaid block in chapter MD; manual override for derived figures.

Anti-patterns

❌ Don't✅ Do
Skip outside-in audit "vì tôi tự biết thesis của mình"3-agent simulation surfaces 30-80 findings inside-out misses
Write thesis in English then translate to VietnameseWrite Vietnamese first; English natural for technical tokens only
Use ## TL;DR sections "vì hiện đại"Strip — academic TÓM TẮT is separate page format
Mention "Claude" / "Wave 12" / "GAP-637" in bodyStrip — bibliography vendor ref [N] Anthropic Claude API only acceptable
Cite internal documents/04-quality/audits/*.md as bibliography refBibliography refs ONLY từ public sources (papers / standards / official docs)
Render Mermaid as code text in DOCXPre-render → PNG via mermaid-cli
110-page bachelor thesis "vì nhiều content"Trim to ≤70 trang — verbose thesis = committee bias against
Skip page-num cite for direct quotes[15, tr.314] mandatory cho direct quote
Use single-child sub-heading (2.1.1 without 2.1.2)Merge content lên parent OR add sibling
Have ## 4.0 Giới thiệu chương + ## Tóm tắt chương 4Drop — fold content into §4.1 + KẾT LUẬN

Related rules + skills

  • .claude/rules/thesis-content-standard.md — 9-category rubric /100 (companion governance)
  • .claude/rules/dev-readable-doc-language.md — Vietnamese narrative + English identifier
  • .claude/rules/output-review-mandate.md §3 — adds row "Thesis report / academic deliverable"
  • .claude/rules/meta-gap-priority.md §3 — META P0 force-multiplier
  • .claude/skills/document-generation/word/ — companion skill for Word/DOCX generation
  • .claude/skills/quality/persona-based-business-review/ — sister skill for 3-agent outside-in audit pattern

Reference

  • reference/outside-in-audit-prompts.md — 3-agent prompt templates (GVHD + GVPB + Defense committee)
  • reference/vn-thesis-conventions.md — UTC/HUST/UET/HCMUT convention notes
  • templates/thesis-info.md.template — student/advisor info template
  • templates/chapter-mapping.md.template — chapter outline template
  • templates/bibliography.md.template — IEEE bibliography template
  • templates/thesis-assemble.py.template — python-docx assembly pipeline template
  • scripts/check-thesis-content-standard.sh — heuristic grep validator for §3 banned patterns

What ships with it: 7 files

39.3 KB alongside SKILL.md, 2 of them executable

Gives 0 of the 12 instructions most literature review skills give in ~4.2k tokens

Counted across 185 of the 191 authors here whose files we hold, read 2026-09-06

  • Screen sources by title, abstract, then full textin 16 of 185, across 9 files
  • Create a search protocol before collecting sourcesin 11 of 185, across 5 files
  • Convert the prompt into a searchable research questionin 11 of 185, across 5 files
  • Keep a reproducible search login 11 of 185, across 5 files
  • Group evidence by theme, not per paperin 11 of 185, across 5 files
  • Ask the user which review rigor level is neededin 11 of 185, across 5 files
  • Use a structured extraction tablein 11 of 185, across 5 files
  • Separate claims by confidence levelin 10 of 185, across 4 files
  • Deduplicate by DOI, identifier, then titlein 10 of 185, across 4 files
  • Record exclusion reasons at each screening stagein 9 of 185, across 3 files
  • Verify citation identifiers before finalizingin 9 of 185, across 4 files
  • Search a minimum of three databasesin 7 of 185, across 6 files

Said here and by no other author read

  • Draft chapter outline before writing chapters
  • Run three-agent outside-in audit before shipping
  • Write chapters in Vietnamese
  • Document measurement methodology for performance claims
  • Define acronyms at first use
  • Score chapters against the rubric

Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.

Keep looking

Skills are one crate of 325,949. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.