DOCX Creator

Thin incremental layer over minimax-skills:minimax-docx. Not a document engine.

Read this first. The OpenXML SDK capability lives in minimax-docx. This skill contains zero engine code. What it contains is the part that took a full debugging session to learn: where that engine's CLI stops being usable, how to drive its SDK correctly for Chinese formal documents, and how to verify the output so you don't ship a file that looks fine to you and broken in Word.

Division of labor

Locate the engine at the marketplace install path, typically ~/.claude/plugins/marketplaces/minimax-skills/skills/minimax-docx/.

Go read minimax-docx directly for anything structural this skill does not cover — images, track changes, comments, TOC, multi-section layouts, template application. Its minimax-docx/references/ folder (cjk_typography.md, openxml_element_order.md, openxml_units.md, troubleshooting.md) and its Samples/*.cs are the authority for SDK patterns. Do not reinvent them here.

Routing: which path for this document?

Rule of thumb: the CLI's --content-json understands exactly three block types (heading, paragraph, pagebreak). Bold, lists, tables, borders, footers, fonts, alignment — none of it is expressible. A Chinese contract needs all of them. See ISSUE-001.

Quick start

The generator reads markdown and writes a formatted .docx. Copy it next to your document so build artifacts stay out of the skill directory:

# 1. Stage the generator beside your markdown
mkdir -p _docxgen && cp <skill-dir>/scripts/Program.cs <skill-dir>/scripts/mmdocx-gen.csproj \
   <skill-dir>/scripts/.gitignore _docxgen/

# 2. Generate (first run restores DocumentFormat.OpenXml + Markdig, ~20s)
dotnet run --project _docxgen -- your-doc.md your-doc.docx

# 3. Structural validation via the engine's XSD validator (note the roll-forward env — ISSUE-003)
DOTNET_ROLL_FORWARD=Major dotnet run \
  --project ~/.claude/plugins/marketplaces/minimax-skills/skills/minimax-docx/scripts/dotnet/MiniMaxAIDocx.Cli \
  -- validate --input your-doc.docx

# 4. MANDATORY visual verification — never skip, never substitute qlmanage
soffice --headless --convert-to pdf --outdir /tmp/docxcheck your-doc.docx
pdftoppm -png -r 100 /tmp/docxcheck/your-doc.pdf /tmp/docxcheck/page
# then Read every /tmp/docxcheck/page-NN.png

# 5. MANDATORY if Microsoft Word.app is installed — LibreOffice cannot see ISSUE-012
open -a "Microsoft Word" your-doc.docx   # check the title bar for "兼容性模式", check every page

Full command details and troubleshooting: scripts/README.md. Full verification steps and pass/fail criteria: references/verification_protocol.md.

Hard rules (violating these means rework)

1. Alignment is layered — this is the expensive one

Never justify the whole document. Three layers, three alignments:

The rule is machine-checkable, so do not eyeball it: if the markdown paragraph's inline tree contains a LineBreakInline, left-align that paragraph; otherwise justify it. Implemented in the case ParagraphBlock p: arm of Main's block-dispatch switch, with the break detection in InlineRuns. (Line numbers are deliberately not given here — they drift every time the file grows; see scripts/README.md's function-name lookup table instead.) Full write-up: ISSUE-004.

2. Every list restarts at 1

Each markdown list must get its own NumId plus a LevelOverride carrying StartOverrideNumberingValue = 1. Reuse one NumId across clauses and clause 3's list silently continues from 4. Implemented across the case ListBlock lb: arm and the NumberingDefinitionsPart setup — see scripts/README.md's lookup table for both. Two SDK traps come with it (wrong class name, wrong element order) — ISSUE-005, ISSUE-006.

3. CJK fonts need both slots

A run must set RunFonts { Ascii, HighAnsi, EastAsia }. Setting only the Latin slots leaves Chinese characters to Word's fallback, and the document renders in whatever the reader's machine picks. Shipped defaults: Latin Times New Roman; East Asian 宋体 for body, 黑体 for headings. Chinese bold runs switch family to 黑体 — 宋体 has no true bold weight and renderer-synthesized bold smears multi-stroke characters (ISSUE-014). Sizes are OpenXML half-points — body 21 (10.5pt, the common Chinese manuscript size), H1 36 (18pt), clause heading 28 (14pt). Body paragraphs also get a 2-character first-line indent (420 twips at 10.5pt — change it in lockstep with the body size). ISSUE-007.

4. Page and table basics

A4 is 11906 × 16838 twips with 1440 twip margins; page number goes in a centered footer PAGE field. Tables need all six borders (top/bottom/left/right/insideH/insideV, in that ECMA-376 order — ISSUE-013) and a tblGrid (ISSUE-012) — set fewer borders and cells look unruled in print; omit tblGrid and the file fails strict validation even though LibreOffice hides it. Implemented in BuildTable and the SectionProperties/FooterPart setup in Main — see scripts/README.md's lookup table.

Verification is not optional

Banned: qlmanage thumbnails as visual proof. macOS Quick Look renders with a different engine than Word and will happily show a clean-looking page for a document whose info blocks are stretched apart. A document was declared "verified perfect" on qlmanage evidence and opened broken in Word — that is the origin of this rule. ISSUE-008.

Required chain: generate → XSD validate → soffice --headless --convert-to pdf → pdftoppm -png → Read every page image → check the five failure modes (info blocks not stretched / each list restarting at 1 / table borders present / signature block intact / no orphaned pagination) → if Microsoft Word.app is installed, open the actual file in it and check the title bar for "兼容性模式" plus every page. Details, prerequisites and the "what counts as failure" list: references/verification_protocol.md.

LibreOffice rendering clean is not the same claim as "Word renders this correctly." ISSUE-012 is a real, reproduced case where a file passed every LibreOffice-based check in this chain — clean info blocks, correct numbering, intact tables — and still opened in real Word with "兼容性模式" in the title bar and phantom bullet markers on paragraphs that were never given any numbering. LibreOffice has no equivalent fallback path and is structurally unable to surface that class of defect. This is ISSUE-008's own lesson (a lenient renderer hides what Word does) recurring one layer up; treat LibreOffice-only verification as incomplete whenever Word is available to check against directly.

Before overwriting a delivered .docx, check for a sibling ~$<name>.docx — that is Word's owner lock, meaning the recipient has the old version open. Overwriting works, but they will keep seeing the stale document until they close and reopen it. Tell them. ISSUE-011.

Customizing

scripts/Program.cs is ~260 lines of straightforward OpenXML. Edit it directly for fonts, sizes, spacing, borders, or new block types — that is the intended workflow. Before adding a structural feature (images, TOC, headers, track changes), read the corresponding Samples/*.cs in minimax-docx first; those patterns are SDK-version-verified and will save you a compile-error loop.

If you append a new property to RunProperties, ParagraphProperties, TableProperties, TableCellProperties, or TableBorders: it must go in ECMA-376's child-element order for that parent, not wherever .Append() reads naturally in the C#. This is not a style preference — get it wrong and you reproduce ISSUE-012/ISSUE-013: minimax-docx validate still reports PASSED, LibreOffice still renders the file cleanly, and real Word still opens it in Compatibility Mode with unrequested formatting scattered across the document. The current code already carries a schema-order comment at each construction site (RunProps, Para, BuildTable) — extend the existing property list in place rather than appending after it, and if you add a genuinely new element, look up its position in openxml_element_order.md (minimax-docx) before deciding where the .Append() call goes. Re-run Step 3a of the verification protocol (real Word, not just LibreOffice) after any such change — that is the only check in the whole chain that would have caught this class of bug.

References

• references/known_issues.md — ISSUE-001…013: symptom / root cause / fix / verification for every trap hit while building this pipeline. Read before debugging anything.
• references/verification_protocol.md — the full end-to-end verification chain, its prerequisites, its pass criteria, and the substitutions that are forbidden.
• scripts/README.md — how to run the generator, what markdown it supports, environment requirements.