kimi-docx
Generate and edit Word documents (.docx). Supports professional documents including covers, charts, track-changes editing, and more. Suitable for any .docx creation or modification task.
By thvroyal · 490 installs
npx skills add thvroyal/kimi-skills --skill kimi-docx
Source repository · Upstream listing
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.
Need Method
Text content only (summarize, analyze, translate) Read tool is fine
Formatting info (copy styles, preserve layout, template filling) Unzip and parse XML
Structure + comments/track changes pandoc input.docx t markdown
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
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:
Fixed Path Conventions (cannot be changed):
Path Purpose
/app/.kimi/skills/kimi docx/ SKILL directory, where commands are executed
/tmp/docx work/ Working directory, edit Program.cs here
/mnt/okcomputer/output/ Output directory, final deliverables
/mnt/okcomputer/upload/ User upload location (input files)
Script Commands ( ./scripts/docx <cmd ):
Command Purpose
env Show environment status (no changes)
init Setup dependencies + workspace
build [out] Compile, run, validate (default: output/output.docx)
validate FILE Validate existing docx
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:
Step Action Notes
1. Compile dotnet build Provides fix suggestions on failure
2. Generate dotnet run <output path Path passed via command line args
3. Auto fix fix element order.py Fixes XML element ordering issues
4. OpenXML validation validator/ Mandatory
5. Business rules validate docx.py Mandatory
6. Statistics Character + word count Optional (requires pandoc)
Validation is mandatory : On failure, file is kept but warnings are shown. Check error messages to fix issues.
Standalone Validation
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:
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
Feature Reason
Watermark Changes visual state. SDK limitation : VML watermark classes don't serialize correctly; must write raw XML to header.
Document protection Restricts editing
Mail merge fields Requires data source
Chart Selection Strategy (Critical)
Default to native Word charts , editable, small file size, professional.
Chart Type Method Notes
Pie chart Native Example.cs → AddPieChart()
Bar chart Native Example.cs → AddBarChart()
Line chart Native Reference bar chart structure, use c:lineChart
Horizontal bar Native Reference bar chart structure, use barDir="bar"
Heatmap, 3D, radar matplotlib Word native doesn't support
Complex statistics (box plot, etc.) matplotlib Word native doesn't support
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() :
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
User Request Execution Standard
Specific word count (e.g., "3000 words") Actual output within ±20%
Specific page count (e.g., "5 pages") Exact match
Range (e.g., "2000 3000 words") Within range
Minimum (e.g., "at least 5000 words") No more than 2x the requirement
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:
Style Palette Suitable Scenarios
Morandi Soft muted tones Artistic, editorial
Earth tones Brown, olive, natural Environmental, organic
Nordic Cool gray, misty blue Minimalist, tech
Japanese Wabi sabi Gray, raw wood, zen Traditional, contemplative
French elegance Off white, dusty pink Luxury, feminine
Industrial Charcoal, rust, concrete Manufacturing, engineering
Academic Navy, burgundy, ivory Research, education
Ocean mist Misty blue, sand Marine, wellness
Forest moss Olive, moss green Nature, sustainability
Desert dusk Ochre, sandy gold Warm, regional
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:
Property XML Effect
Keep with next <w:keepNext/ Heading stays on same page as following paragraph
Keep lines together <w:keepLines/ Paragraph won't break across pages
Page break before <w:pageBreakBefore/ Force new page (for H1)
Widow/orphan control <w:widowControl/ Prevent single lines at top/bottom of page
Table pagination :
Part 4: Technical Reference
Choose your path:
Task Stack Reference
Create new document C + OpenXML SDK 4.1 4.6 + Example.cs
Edit existing document Python + lxml 4.7 + references/EditingGuide.md
4.1 SDK Fundamentals
Schema Compliance (MEMORIZE THESE)
OpenXML has strict element ordering requirements. Wrong order = Word cannot open the file.
Required Styles
Element Order Rules
Most ordering issues are auto fixed by fix element order.py . Key rules to remember:
Parent Key Rule
sectPr headerRef → footerRef must come before pgSz → pgMar
Table Must have tblGrid between tblPr and tr (see below)
Tables Must Have tblGrid
Table Column Width Consistency
Main cause of skewed tables: gridCol width in tblGrid doesn't match cell's tcW width.
Rule Reason
gridCol count = table column count Otherwise column width calculation fails
gridCol.Width = tcW.Width Mismatch causes skewing (checked during validation)
All rows in same column use same tcW Maintains column width consistency
Value Limits
paraId must be < 0x80000000 (for comment paragraph IDs)
Creation vs Editing
Task Method Why
Create new document C OpenXML SDK Handles package structure, rels, Content Types automatically
Edit existing document Python + lxml Transparent, no black box, full control
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 overa