Word Typesetting

A reusable Codex / Claude Code skill that formats formal Chinese Word (.docx) reports to a consistent academic standard — course designs, internship reports, thesis-style documents, and engineering calculation reports.

It is two things working together:

1. A written standard (SKILL.md + references/) that tells the model exactly how a compliant report should look — typography, document structure, tables, figures, formulas, citations, numbers/units.
2. Tooling (scripts/) that checks a finished .docx against the machine-checkable parts of that standard, and auto-fixes the safest issues.

The model does the actual editing (with the general DOCX tooling); this skill supplies the rules, a completion checklist, and a guardrail so nothing obvious slips through.

---

What it can do

Typography & language

• Chinese body text in Songti/SimSun, English/numbers in Times New Roman, headings in Heiti.
• A full font-size hierarchy (title 小二, H1 三号, H2 四号, H3 小四, body 小四, captions 五号, headers/footers/footnotes 小五).
• Dominant-language punctuation normalization (Chinese vs. English), special marks (……, ——, 《 》), while protecting URLs, DOIs, code, formulas, decimals, and citation brackets.
• Body justified, first-line indent 2 字符, 1.5× or fixed 20–22 pt line spacing.

Document structure & page setup

• Standard section order (cover → declarations → abstracts → TOC → body → references → appendix → acknowledgements).
• A4 page setup, margins/gutter, headers/footers, Roman→Arabic page numbering via section breaks.
• Auto-generated table of contents (with optional figure/table lists), multilevel heading numbering (1 / 1.1 / 1.1.1), widow/orphan control.
• Headings left-aligned (title centered), no trailing punctuation, depth ≤ 3–4 levels.

Tables, figures, formulas

• White three-line tables (top/bottom rules ~1.5 pt, header rule ~0.75 pt), body small-four/12 pt and centered, caption above the table, header row repeated across page breaks.
• Figures with caption below, by-chapter numbering (图 1-1 / 表 2-3 / (3-1)), the "reference-before-appearance" rule.
• Formulas authored as LaTeX and rendered to native Word OMML — variables italic, units/operators/functions upright; equation numbers right-aligned per chapter, cited as "由式 (3-1) 可得".

Citations & references

• In-text citations as superscript, field-backed cross-references to the whole bracketed [1] (brackets included, not a bare number).
• Bibliography entries per GB/T 7714—2015 (numeric sequential-coding by default; author-year supported), with the right type tags ([J] [M] [D] [C] [S] [P] [EB/OL] …).

Numbers, units, appendix code

• Space between number and unit (20 m³/s), GB 3100/3101/3102 units, GB/T 15835 numeral usage.
• Appendix code after the main text with red comments (C00000).

What the result looks like

A polished, navigable report: a clickable heading outline and an auto-updating TOC; left-aligned Heiti headings over Songti body text at a consistent size; clean white three-line tables with captions above and figures with captions below, all numbered by chapter; real Word equation objects instead of pasted images or plain text; superscript [1]-style citations that renumber themselves because they are Word cross-reference fields; and a GB/T 7714—2015 reference list. Running the auditor on a finished file prints PASS: no machine-detected guardrail issues.

See examples/ for a deliberately non-compliant sample and the audit report it produces — a concrete before/after of what the rules catch.

Install

Install from GitHub:

AllenWang2005/Word-typesetting

Or copy the folder into your skills directory:

~/.codex/skills/word-report-formatting                 # Codex, macOS / Linux
%USERPROFILE%\.codex\skills\word-report-formatting     # Codex, Windows
~/.claude/skills/word-report-formatting                # Claude Code

Use

Ask the assistant to use the skill when creating or polishing a Word report:

Use $word-report-formatting to format this course design report.

It also triggers naturally for formal Chinese Word reports, three-line tables, OMML formulas, figure/table captions, citation cross-references, or appendix code.

The audit script

After formatting a DOCX, run the guardrail:

python scripts/audit_docx_format.py path/to/report.docx
python scripts/audit_docx_format.py path/to/report.docx --json   # machine-readable

It reports FAIL (machine-certain violations) and WARN (needs a human look), ends with a FAIL/WARN summary, and notes any per-code truncation. Checks include:

Code	Severity	What it catches
ZH_PUNCT / EN_PUNCT	FAIL	Wrong-language punctuation in prose
ABSTRACT_INDENT	FAIL	Abstract/keywords centered or indented
H1_CENTER	FAIL	Level-1 heading centered instead of left-aligned
H1_GAP	WARN	Level-1 heading missing a blank gap before it
HEADING_PUNCT	WARN	Heading ends with punctuation
HEADING_FONT / BODY_FONT	WARN	Heading in Songti / body in Heiti (direct fonts only)
TABLE_SIZE	FAIL	Table text not small-four/12 pt
CAPTION_POSITION	WARN	Table caption below table / figure caption above figure
COLOR	WARN	Stray non-black font color (hyperlinks/theme colors excluded)
CITATION_BRACKETS	FAIL	Full-width citation brackets ［1］
CITATION_FIELDS	FAIL	Citations exist but no REF ref_### fields
CITATION_NO_BRACKETS	WARN	Bare superscript number citation missing its brackets
VISIBLE_LATEX	FAIL	Visible LaTeX source instead of OMML
FORMULA_TEXT	FAIL/WARN	Plain-text formula / quantity symbol that should be OMML

It is intentionally domain-neutral — no hard-coded field-specific symbol list.

Audit scope: the script inspects the main document story in word/document.xml. It does not audit headers, footers, footnotes, endnotes, comments, or embedded parts — verify those visually or with a deeper OOXML pass when they matter.

The auto-fix script

scripts/normalize_docx.py mechanically fixes the two safest issues — full-width citation brackets (［1］ → [1]) and ASCII sentence punctuation between CJK characters (中文,中文 → 中文，中文) — preserving every other byte of the package:

python scripts/normalize_docx.py report.docx -o report.fixed.docx
python scripts/normalize_docx.py report.docx --in-place

It deliberately does not touch fonts, styles, formulas, or cross-references; those need judgement and stay with the model + the main standard.

Tests

Standard library only (no third-party dependencies):

python -m unittest discover -s tests -v

CI runs py_compile plus the test suite on Python 3.9 and 3.12 on every push (see .github/workflows/ci.yml).

Repository structure

.
├── SKILL.md
├── LICENSE
├── agents/
│   └── openai.yaml
├── references/
│   ├── formatting-standard.md                  # main checklist
│   ├── document-structure-and-page-setup.md    # structure, page setup, TOC, font sizes
│   ├── latex-omml-formula-workflow.md          # LaTeX → Word OMML workflow
│   ├── reference-style-gbt7714.md              # GB/T 7714—2015 bibliography format
│   └── citation-crossrefs-ooxml.md             # in-text REF cross-reference OOXML
├── scripts/
│   ├── audit_docx_format.py                    # read-only guardrail
│   └── normalize_docx.py                       # safe auto-fixer
├── examples/
│   ├── README.md
│   ├── make_sample.py
│   └── sample-audit-output.txt
└── tests/
    ├── test_audit_docx_format.py
    └── test_normalize_docx.py

Maintenance notes

• Keep SKILL.md concise so it loads quickly when the skill triggers.
• Put detailed rules in references/.
• Keep the auditor conservative: FAIL only for machine-checkable violations, WARN for items needing visual review.
• Do not store private information, credentials, or one-off chat history in this repository.
• When the standard changes, update the relevant reference file and refresh the summary/paths in SKILL.md.

License

MIT.