Kimi DOCX

Part 1: Goals

⚠️ When to Unzip vs Read

To preserve ANY formatting from the source document, MUST unzip and parse XML.

Read tool returns plain text only — fonts, colors, alignment, borders, styles are lost.

Core Principles

1. Preserve formatting — When editing existing documents, retain original formatting. Clone and modify, never recreate.

2. Correct feature implementation — Comments need multi-file sync. Track Changes need revision marks. Use the right structure.

Never use python-docx/docx-js as fallback. These libraries produce lower quality output than direct XML manipulation.

Source Principle

Template provided = Act as form-filler, not designer.

• Format is the user's decision
• Task: replace placeholders, not redesign
• Like filling a PDF form—do not redesign

No template = Act as designer. Design freely based on scenario.

For .doc (legacy format), first convert with libreoffice --headless --convert-to docx.

Part 2: Execution

File Structure

docx/
├── SKILL.md                      ← This file (entry point + reference)
├── references/
│   └── EditingGuide.md           → Complete Python editing tutorial (comments, track changes, 5-file sync)
├── scripts/
│   ├── docx                      → Unified entry (the only script to call)
│   ├── fix_element_order.py      → Auto-fix XML element ordering
│   ├── validate_docx.py          → Business rule validation
│   ├── generate_backgrounds.py       → Morandi style backgrounds
│   ├── generate_inkwash_backgrounds.py → Ink-wash style backgrounds
│   └── generate_chart.py         → matplotlib (only for heatmaps/3D/radar; simple charts must use native)
├── assets/
│   └── templates/
│       ├── KimiDocx.csproj       → Project file template (for creating new docs)
│       ├── Program.cs            → Program entry template
│       ├── Example.cs            → Complete example (cover+TOC+charts+back cover)
│       ├── CJKExample.cs         → CJK content patterns (quote escaping, fonts)
│       └── xml/                  → XML templates for comments infrastructure
└── validator/                    → OpenXML validator (pre-compiled binary, AI does not modify)

Creating new documents: Use C# SDK with ./scripts/docx build → See Example.cs for patterns, CJKExample.cs for CJK content Editing existing documents: Use Python + lxml → See references/EditingGuide.md for complete tutorial

⚠️ Do NOT mix these approaches. C# SDK for creation, Python for editing. Never use python-docx/docx-js.

Environment Setup

First time, execute in the SKILL directory:

cd /app/.kimi/skills/kimi-docx/
./scripts/docx init

Fixed Path Conventions (cannot be changed):

Script Commands (./scripts/docx <cmd>):

The script automatically handles:

• Detects dotnet/python3 (required), pandoc/playwright/matplotlib (optional)
• Installed → use directly; Not installed → auto-install; Broken → repair
• Initializes working directory, copies template files

Build Process

Must use ./scripts/docx build, do not execute dotnet build && dotnet run separately (skips validation).

Program.cs Output Path Convention (Critical)

Program.cs must get output path from command line arguments, otherwise build script cannot find the generated file:

// Correct - get output path from command line arguments
string outputFile = args.Length > 0 ? args[0] : "/mnt/okcomputer/output/output.docx";

// Wrong - hardcoded path causes build failure
string outputFile = "my_document.docx";  // Script can't find file!

Validation is mandatory: On failure, file is kept but warnings are shown. Check error messages to fix issues.

Standalone Validation

cd /app/.kimi/skills/kimi-docx/
./scripts/docx validate /mnt/okcomputer/output/report.docx

Content Verification (Mandatory)

pandoc is the SOURCE OF TRUTH. OpenXML validator checks structure; pandoc shows actual content.

Before delivery, verify with pandoc:

• pandoc output.docx -t plain — check text completeness
• For revisions/comments: add --track-changes=all to verify marker positions

⚠️ Critical: comments.xml exists ≠ comments visible. Count mismatch = doc_tree not saved. See references/EditingGuide.md §5.3.

Part 3: Quality Standards

Delivery Standard

Generic styling and mediocre aesthetics = mediocre delivery.

Deliver studio-quality Word documents with deep thought on content, functionality, and styling. Users often don't explicitly request advanced features (covers, TOC, backgrounds, back covers, footnotes, charts)—deeply understand needs and proactively extend.

Language Consistency

Document language = User conversation language (including filename, body text, headings, headers, TOC hints, chart labels, and all other text).

Headers and Footers - REQUIRED BY DEFAULT

Most documents MUST include headers and footers. The specific style (alignment, format, content) should match the document's overall design.

• Header: Typically document title, company name, or chapter name
• Footer: Typically page numbers (format flexible: "X / Y", "Page X", "— X —", etc.)
• Cover/Back cover: Use TitlePage setting to hide header/footer on first page

Professional Elements (Critical)

Create documents that exceed user expectations, proactively add professional elements, don't wait for users to ask. Delivery standard: Visual quality of a top designer in 2024.

Cover & Visual:

• Formal documents (proposals, reports, financials, bids, contracts) / creative documents (invitations, greeting cards) must have cover and back cover
• Covers must have designer-quality background images
• Body pages can optionally include backgrounds to enhance visual appeal

Structure:

• Long documents (3+ sections) add TOC, must add refresh hint after TOC

Data Presentation:

• When comparing data or showing trends, use charts instead of plain text lists
• Tables use light gray headers or three-line style, avoid Word default blue

Links & References:

• URLs must be clickable hyperlinks
• Multiple figures/tables add numbering and cross-references ("see Figure 1", "as shown in Table 2")
• Academic/legal/data analysis citation scenarios implement correct in-text click-to-jump references with corresponding footnotes/endnotes

TOC Refresh Hint

Word TOC is field code, page numbers may be inaccurate when generated. Must add gray hint text after TOC, informing users to manually refresh:

Table of Contents
─────────────────
Chapter 1 Overview .......................... 1
Chapter 2 Methods ........................... 3
...

(Hint: On first open, right-click the TOC and select "Update Field" to show correct page numbers)

Hint text requirements:

• Visually subtle — gray color, smaller font size, should not compete with actual TOC entries
• Language: Matches user conversation language

Only When User Explicitly Requests

Chart Selection Strategy (Critical)

Default to native Word charts, editable, small file size, professional.

Native charts are preferred (editable, smaller files), but matplotlib is acceptable for data analysis scenarios.

Inserting Images/Charts

Any PNG (matplotlib charts, backgrounds, photos) must be inserted using AddInlineImage():

AddInlineImage(body, mainPart, "/path/to/image.png", "Description", docPrId++);

Critical:

• Chart labels/titles must match document language (e.g., Chinese labels for Chinese docs)
• Build output shows X images — if 0, images were not inserted

Content Constraints

Word/Page Count Requirements

Forbidden: Padding word count with excessive bullet point lists. Maintain information density.

Outline Adherence

• User provides outline: Follow strictly, no additions, deletions, or reordering
• No outline provided: Use standard structure
  • Academic: Introduction → Literature → Methods → Results → Discussion → Conclusion
  • Business: Executive Summary → Analysis → Recommendations
  • Technical: Overview → Principles → Usage → Examples → FAQ

Scene Completeness

Think one step ahead of the user, complete elements the scenario needs. Examples below are not exhaustive — apply this principle to ALL document types:

• Exam paper → Name/class/ID fill areas, point allocation per question (consider total), grading section
• Contract → Signature and seal areas for both parties, date, contract number, attachment list
• Meeting minutes → Attendees, absentees, action items with owners, next meeting time

Design Philosophy

Color Scheme

Low saturation tones, avoid Word default blue and matplotlib default high saturation.

Flexibly choose color schemes based on document scenario:

Color scheme must be consistent within the same document.

Layout

White space (margins, paragraph spacing), clear hierarchy (H1 > H2 > body), proper padding (text shouldn't touch borders).

Pagination Control

Word uses flow layout, not fixed pages. Control pagination with these properties:

// Example: H1 always starts on new page, stays with next paragraph
new ParagraphProperties(
    new ParagraphStyleId { Val = "Heading1" },
    new PageBreakBefore(),
    new KeepNext(),
    new KeepLines()
)

Table pagination:

// Allow row to break across pages (avoid large blank areas)
new TableRowProperties(
    new CantSplit { Val = false }  // false = can split
)

// Repeat header row on each page
new TableRowProperties(
    new TableHeader()
)

Part 4: Technical Reference

Choose your path:

4.1 SDK Fundamentals

Schema Compliance (MEMORIZE THESE)

OpenXML has strict element ordering requirements. Wrong order = Word cannot open the file.

Required Styles

// Normal style must exist - all Heading styles use basedOn="Normal"
styles.Append(new Style(
    new StyleName { Val = "Normal" },
    new StyleParagraphProperties(
        new SpacingBetweenLines { After = "200", Line = "276", LineRule = LineSpacingRuleValues.Auto }
    ),
    new StyleRunProperties(
        new RunFonts { Ascii = "Calibri", HighAnsi = "Calibri" },
        new FontSize { Val = "22" },
        new FontSizeComplexScript { Val = "22" }
    )
) { Type = StyleValues.Paragraph, StyleId = "Normal", Default = true });

Element Order Rules

Most ordering issues are auto-fixed by fix_element_order.py. Key rules to remember:

Tables Must Have tblGrid

// Correct - table must define grid
var table = new Table();
table.Append(new TableProperties(...));
table.Append(new TableGrid(           // Required!
    new GridColumn { Width = "4680" },
    new GridColumn { Width = "4680" }
));
table.Append(new TableRow(...));

// Wrong - missing tblGrid, Word cannot open
var table = new Table();
table.Append(new TableProperties(...));
table.Append(new TableRow(...));  // Adding rows directly

Table Column Width Consistency

Main cause of skewed tables: gridCol width in tblGrid doesn't match cell's tcW width.

// Correct - gridCol and tcW match exactly
table.Append(new TableGrid(
    new GridColumn { Width = "3600" },  // First column
    new GridColumn { Width = "5400" }   // Second column
));

var row = new TableRow(
    new TableCell(
        new TableCellProperties(
            new TableCellWidth { Width = "3600", Type = TableWidthUnitValues.Dxa }  // Matches gridCol!
        ),
        new Paragraph(new Run(new Text("Content")))
    ),
    new TableCell(
        new TableCellProperties(
            new TableCellWidth { Width = "5400", Type = TableWidthUnitValues.Dxa }  // Matches gridCol!
        ),
        new Paragraph(new Run(new Text("Content")))
    )
);

Value Limits

• paraId must be < 0x80000000 (for comment paragraph IDs)

Creation vs Editing

For creating new documents: Use Example.cs patterns with SDK.

For editing existing documents: See references/EditingGuide.md for complete Python workflow.

Example.cs

Read the entire file to understand the overall structure, not just individual functions. The file demonstrates how sections connect (cover → TOC → body → back cover).

The "Project Proposal", "[Company Name]", etc. in Example are example content only, and the color scheme is for reference only.

⚠️ Do NOT copy the Example's color scheme. Redesign visual style based on YOUR document's scenario, like a top designer.

Function Index (read source for implementation details):

CJKExample.cs

CJK documents must read CJKExample.cs only — reading Example.cs instead will cause errors (missing font config, quote escaping). It handles:

• Quote escaping ("" → \u201c \u201d)
• CJK font configuration (SimHei, Microsoft YaHei)
• Paragraph indentation for CJK text

Structure is identical to Example.cs — no need to read both.

4.2 Content Elements

Field Codes

PAGE/NUMPAGES/DATE/TOC — structure: FieldChar(Begin) → FieldCode(" PAGE ") → FieldChar(Separate) → Text → FieldChar(End). Results cached; WPS doesn't support UpdateFieldsOnOpen.

Bookmarks and Cross-References

Bookmarks mark positions (BookmarkStart/BookmarkEnd with matching IDs); cross-references link via REF field (" REF bookmarkName \\h ").

Pitfall: Deleting bookmarked text deletes bookmark → "Error! Reference source not found".

4.3 Visual Design

Background Image Design

Cover/back cover must have background. Background images should have center white space, use low saturation colors. Background images must NOT contain any text; text should be implemented in Word for user editability.

Design Flow

1. Read example: Read scripts/generate_backgrounds.py for HTML/CSS techniques (radial-gradient, transparency, positioning)
2. Choose direction: Select a style direction from the table below based on document scenario
3. Create original: Write new HTML/CSS from scratch—the example shows ONE style, yours should be different

⚠️ Copying the example = all documents look the same = mediocre delivery. Each document deserves a unique visual identity matching its content and purpose.

Style Reference

Technical: Playwright generates 794×1123px (device_scale_factor=2), insert as floating Anchor with BehindDoc=true. See Example.cs:CreateFloatingBackground().

Letterhead (Business Documents)

For formal business letters, consider adding a letterhead in the header area. Common patterns:

• Full letterhead on first page (logo + company name + contact info), simplified or hidden on subsequent pages
• Use TitlePage in SectionProperties to enable different first-page header
• Design flexibly based on the specific business context—no fixed rules

Two-Column Layout

Use sectPr with Columns. Affects entire section until next sectPr.

4.4 Special Content

Math Formulas (OMML)

Core pattern: <m:e> is the universal content container. Almost all elements wrap content in <m:e>.

Text: Always <m:r><m:t>text</m:t></m:r>, never bare text.

Root: <m:oMath> (inline) or <m:oMathPara> (display). Do NOT nest <m:oMath> inside another.

Structure examples:

Trap: Matrix uses <m:e> for cells, NOT <m:mc> (which is for column properties).

Curly Quotes in C# Strings

C# treats " " as string delimiters → CS1003. Simplest fix: Use escaped straight quotes \" in string literals. If curly quotes are required, use XML entity encoding: “ ” (doubles) or ‘ ’ (singles).

Chinese quote handling — see CJKExample.cs for complete patterns:

// ❌ Wrong - Chinese quotes break compilation
new Text("请点击"确定"按钮")  // CS1003!

// ✓ Correct - use Unicode escapes
new Text("请点击\u201c确定\u201d按钮")

⚠️ Do NOT use verbatim strings @"" — \u escapes don't work in verbatim strings:

// ❌ WRONG - @"" verbatim string, \u NOT escaped, outputs literal "\u201c"
string text = @"她说\u201c你好\u201d";  // Outputs: 她说\u201c你好\u201d

// ✓ CORRECT - regular string, \u IS escaped
string text = "她说\u201c你好\u201d";   // Outputs: 她说"你好"

// ✓ For long text, use + concatenation
string para = "第一段内容，" +
              "她说\u201c这是引用\u201d，" +
              "继续写第二段。";

Units

Twips = 1/20 pt (11906 = A4 width). Half-points for font size (24 = 12pt). EMU = 914400/inch.

4.5 Page Layout

Image Size

wp:extent and a:ext Cx/Cy must match. For proportional scaling: read PNG header (bytes 16-23) for dimensions, calculate cy = cx * height / width.

Pagination Control

Add KeepNext to title/chart paragraphs to prevent orphaned titles or chart-caption separation.

Section Breaks

sectPr inside pPr = last paragraph of section. Avoid PageBreak + Continuous (blank page). Use NextPage.

Table of Contents (TOC)

WPS doesn't support UpdateFieldsOnOpen → must pre-populate TOC entries using field code structure: FieldChar(Begin) → FieldCode(" TOC ...") → FieldChar(Separate) → placeholder entries (hyperlinked text + page numbers) → FieldChar(End). The placeholder entries between Separate and End allow Word to display a TOC immediately; users refresh to get accurate page numbers. Never use static text paragraphs to simulate a TOC—must use field code structure, otherwise it cannot be refreshed. See Example.cs:AddTocSection().

Parameters: \o "1-3" (heading levels), \h (hyperlinks), \z (hide page# in web), \u (outline level).

Headings must use built-in Heading1/Heading2 styles (custom styles not recognized).

Alignment and Typography

CJK body: justify + 2-char indent. English: left. Table numbers: right. Headings: no indent.

4.6 Page Elements

Headers and Footers

// 1. Create header part
var headerPart = mainPart.AddNewPart<HeaderPart>();
var headerId = mainPart.GetIdOfPart(headerPart);

headerPart.Header = new Header(
    new Paragraph(
        new ParagraphProperties(
            new ParagraphStyleId { Val = "Header" },
            new Justification { Val = JustificationValues.Center }
        ),
        new Run(new Text("Document Title"))
    )
);

// 2. Create footer part (with page numbers)
var footerPart = mainPart.AddNewPart<FooterPart>();
var footerId = mainPart.GetIdOfPart(footerPart);

var footerPara = new Paragraph(
    new ParagraphProperties(
        new Justification { Val = JustificationValues.Center }
    )
);
// PAGE field: Begin → FieldCode → Separate → Text → End
footerPara.Append(new Run(new FieldChar { FieldCharType = FieldCharValues.Begin }));
footerPara.Append(new Run(new FieldCode(" PAGE ")));
footerPara.Append(new Run(new FieldChar { FieldCharType = FieldCharValues.Separate }));
footerPara.Append(new Run(new Text("1")));  // Placeholder, updated on open
footerPara.Append(new Run(new FieldChar { FieldCharType = FieldCharValues.End }));
footerPara.Append(new Run(new Text(" / ") { Space = SpaceProcessingModeValues.Preserve }));
// NUMPAGES field (same structure)
footerPara.Append(new Run(new FieldChar { FieldCharType = FieldCharValues.Begin }));
footerPara.Append(new Run(new FieldCode(" NUMPAGES ")));
footerPara.Append(new Run(new FieldChar { FieldCharType = FieldCharValues.Separate }));
footerPara.Append(new Run(new Text("1")));
footerPara.Append(new Run(new FieldChar { FieldCharType = FieldCharValues.End }));
footerPart.Footer = new Footer(footerPara);

// 3. Reference in SectionProperties
new SectionProperties(
    new HeaderReference { Type = HeaderFooterValues.Default, Id = headerId },
    new FooterReference { Type = HeaderFooterValues.Default, Id = footerId },
    new PageSize { Width = 11906, Height = 16838 },
    new PageMargin { Top = 1440, Right = 1440, Bottom = 1440, Left = 1440, Header = 720, Footer = 720 }
)

Header/Footer Types:

Different first page (for cover): add TitlePage() to sectPr.

Different odd/even pages: add <w:evenAndOddHeaders/> in settings.xml.

Footnotes and Endnotes

Separator trap: FootnotesPart/EndnotesPart must include Id=-1 (Separator) and Id=0 (ContinuationSeparator) before any user notes. Missing these → Word fails to render.

<!-- Required in footnotes.xml / endnotes.xml before user notes -->
<w:footnote w:type="separator" w:id="-1">
  <w:p><w:r><w:separator/></w:r></w:p>
</w:footnote>
<w:footnote w:type="continuationSeparator" w:id="0">
  <w:p><w:r><w:continuationSeparator/></w:r></w:p>
</w:footnote>
<!-- User notes start from id="1" -->

Lists

Requires NumberingDefinitionsPart with AbstractNum + NumberingInstance. Apply via NumberingProperties in paragraph.

Multi-level: create AbstractNum with multiple Levels. Formats: Decimal, UpperLetter, LowerRoman, Bullet, ChineseCounting.

Hyperlinks

Must use <w:hyperlink> element, not plain text. Requires relationship first:

var relId = mainPart.AddHyperlinkRelationship(new Uri("https://example.com"), true).Id;
paragraph.Append(new Hyperlink(new Run(
    new RunProperties(new Color { Val = "0563C1" }, new Underline { Val = UnderlineValues.Single }),
    new Text("Click here")
)) { Id = relId })

Charts and Visualization

Word Chart: Use NumberLiteral (no Excel), DataPoint for colors. See Example.

matplotlib: dpi=300, axes.unicode_minus=False. Font/labels must match document language.

4.7 Editing Operations (Python API)

Use docx_lib.editing for comments and track changes:

from scripts.docx_lib.editing import (
    DocxContext,
    add_comment, reply_comment, resolve_comment, delete_comment,
    insert_paragraph, insert_text, propose_deletion,
    reject_insertion, restore_deletion, enable_track_changes
)

with DocxContext("input.docx", "output.docx") as ctx:
    # add_comment(ctx, para_text, comment, highlight=None)
    # - para_text: text to locate paragraph
    # - comment: comment content
    # - highlight: text to highlight (omit to highlight entire paragraph)
    add_comment(ctx, "M-SVI index", "Please define", highlight="M-SVI")
    insert_text(ctx, "The method", after="method", new_text=" and materials")

Complete guide: references/EditingGuide.md

4.8 XML Quick Reference

Text Formatting (rPr)

<w:r>
  <w:rPr>
    <w:rFonts w:ascii="Times New Roman" w:eastAsia="SimSun"/>
    <w:sz w:val="24"/>  <!-- 12pt = 24 half-points -->
    <w:b/><w:i/><w:u w:val="single"/>
    <w:color w:val="FF0000"/>
  </w:rPr>
  <w:t>text</w:t>
</w:r>

Font sizes: 21=10.5pt, 24=12pt, 28=14pt, 32=16pt, 44=22pt

Track Changes Structure

<!-- Insertion: <w:ins> wraps <w:r> -->
<w:ins w:id="1" w:author="..." w:date="...">
  <w:r><w:rPr>...</w:rPr><w:t>text</w:t></w:r>
</w:ins>

<!-- Deletion: <w:del> wraps <w:r> (same pattern as ins!) -->
<w:del w:id="2" w:author="..." w:date="...">
  <w:r><w:rPr>...</w:rPr><w:delText>text</w:delText></w:r>
</w:del>

Key: Both <w:ins> and <w:del> wrap <w:r>, not inside it. Use <w:delText> instead of <w:t> for deletions.

Schema Constraints

Complete examples: See references/EditingGuide.md for full working code.