DOCX Skill

DOCX Skill

Read with WordInspect; all write / modify / create operations use JsSandbox to run Node scripts (preferred library: docx); format conversion uses DocConvert.

Tool List

Loading (two steps):

1. LoadSkill docx-skill
2. ToolSearch(query: "select:WordInspect,JsSandbox,DocConvert")

Read / DocPreview on a .docx returns only Markdown-level body text (losing rPr / pPr / table merges / tracked changes / comments); any "analyze / summarize / edit / create / review Word" task must go through WordInspect + JsSandbox.

1. Read: Start with WordInspect(summary)

WordInspect { action: "summary", filePath: "…" }

Returns pageCount / wordCount / headingCount / hasTrackedChanges / hasComments / isProtected / availableStyles. Branch accordingly:

2. Write: JsSandbox + docx

Use the docx npm package (flumens/docx) to programmatically build docx OOXML. Broadest coverage for styles / tracked changes / comments.

2.1 Demo: Generate a meeting notes document

import {
  Document, Packer, Paragraph, TextRun, HeadingLevel, Table, TableRow, TableCell,
  WidthType, AlignmentType, PageOrientation,
} from 'docx'
import fs from 'node:fs/promises'

const title = new Paragraph({
  heading: HeadingLevel.HEADING_1,
  alignment: AlignmentType.CENTER,
  children: [new TextRun({ text: 'Q2 Product Planning — Meeting Notes', bold: true, size: 36 })],
})

const meta = new Paragraph({
  children: [new TextRun({ text: 'Date: 2026-04-20   Attendees: Alice / Bob / Carol', size: 22 })],
})

const decisions = [
  ['Alice', 'Q2 feature schedule', '2026-05-10'],
  ['Bob', 'API documentation', '2026-04-30'],
]
const tbl = new Table({
  width: { size: 100, type: WidthType.PERCENTAGE },
  rows: [
    new TableRow({
      tableHeader: true,
      children: ['Owner', 'Item', 'Due Date'].map(
        t => new TableCell({
          children: [new Paragraph({ children: [new TextRun({ text: t, bold: true })] })],
        }),
      ),
    }),
    ...decisions.map(row => new TableRow({
      children: row.map(v => new TableCell({ children: [new Paragraph(v)] })),
    })),
  ],
})

const doc = new Document({
  creator: 'OpenLoaf',
  styles: {
    default: {
      document: { run: { font: 'Calibri' } },
    },
  },
  sections: [{
    properties: { page: { orientation: PageOrientation.PORTRAIT } },
    children: [title, meta, new Paragraph({ text: '' }), tbl],
  }],
})

await fs.writeFile('meeting_notes.docx', await Packer.toBuffer(doc))
console.log('meeting_notes.docx written')

CJK font: set styles.default.document.run.font to 'Microsoft YaHei' / 'SimSun' / 'Noto Sans CJK SC' to avoid rendering squares in Word.

2.2 Demo: Text replacement (simple find-replace)

In docx, {{placeholder}} tokens may be split across multiple w:r runs by Word. Use WordInspect(xml) to check the structure. When the placeholder lives in a single run:

import AdmZip from 'adm-zip'
import fs from 'node:fs/promises'

const zip = new AdmZip(await fs.readFile('in.docx'))
let xml = zip.readAsText('word/document.xml')
xml = xml
  .replace(/\{\{date\}\}/g, '2026-04-20')
  .replace(/\{\{name\}\}/g, 'John Smith')
zip.updateFile('word/document.xml', Buffer.from(xml, 'utf-8'))
await fs.writeFile('out.docx', zip.toBuffer())
console.log('replaced')

When runs are split, rewriting the whole new Document({...}) is more reliable.

2.3 Demo: Insert image + header logo

import {
  Document, Packer, Paragraph, ImageRun, Header, AlignmentType,
} from 'docx'
import fs from 'node:fs/promises'

const logo = await fs.readFile('logo.png')

const doc = new Document({
  sections: [{
    headers: {
      default: new Header({
        children: [new Paragraph({
          alignment: AlignmentType.RIGHT,
          children: [new ImageRun({
            data: logo,
            transformation: { width: 80, height: 24 },
          })],
        })],
      }),
    },
    children: [
      new Paragraph({ text: 'Report body...' }),
      new Paragraph({
        children: [new ImageRun({
          data: await fs.readFile('chart.png'),
          transformation: { width: 500, height: 300 },
        })],
      }),
    ],
  }],
})
await fs.writeFile('report.docx', await Packer.toBuffer(doc))

2.4 Demo: Accept all tracked changes

No direct API; use adm-zip to edit document.xml:

import AdmZip from 'adm-zip'
import fs from 'node:fs/promises'

const zip = new AdmZip(await fs.readFile('in.docx'))
let xml = zip.readAsText('word/document.xml')
// 1) Accept all w:ins (keep content, remove tags)
xml = xml.replace(/<w:ins [^>]*>([\s\S]*?)<\/w:ins>/g, '$1')
// 2) Accept all w:del (remove the deleted content)
xml = xml.replace(/<w:del [^>]*>[\s\S]*?<\/w:del>/g, '')
zip.updateFile('word/document.xml', Buffer.from(xml, 'utf-8'))
await fs.writeFile('out.docx', zip.toBuffer())
console.log('tracked changes accepted')

3. Format Conversion with DocConvert

DocConvert(from="docx", to="pdf",  sourcePath="…")
DocConvert(from="docx", to="md",   sourcePath="…")
DocConvert(from="docx", to="html", sourcePath="…")

Backed by LibreOffice headless; layout fidelity is superior to hand-coded conversion.

4. Common Pitfalls

To patch a script: JsSandbox(action="edit-and-run", scriptPath=<previous path>, edits=[{find,replace}]) — only pass the diff.