Best Claude DOCX Skills: Word Skills for AI Coding Agents

If your Word documents contain tracked changes, comments or a company template, the skill you pick matters more than the model behind it. A Claude DOCX skill decides whether the agent edits the XML inside the file, rebuilds the document from scratch or runs a script that keeps your styles intact. This guide compares the skills in the directory's docx topic and says which job each one suits.

Facts come from each skill's SKILL.md and repository. The editorial team did not run them on sample documents, so nothing here rates output quality. For every pick you get a "best for" line, requirements, the licence as the directory records it, the automated safety result in plain words, and a trade-off.

The main ways a DOCX skill works

A .docx file is a ZIP archive of XML parts, and the skills divide into three approaches.

• docx-js plus XML edits. New documents are written in JavaScript with the docx package, and existing files are unzipped, edited and repacked. The official skill and two close relatives work this way.
• python-docx plus raw XML for the hard parts. Creating and simple edits use python-docx, and features it cannot express, such as tracked changes and comments, drop to raw XML.
• Scripted operations. The agent runs bundled scripts that inspect, replace or create content, which makes the outcome easier to repeat.

Which approach fits depends on the languages already in your environment and on how much you care about preserving an existing document byte for byte.

Claude DOCX skills at a glance

The official docx skill from Anthropic

Best for: producing a formatted report, memo or letter, and editing an existing .docx without breaking its formatting.

anthropics/docx is in Anthropic's skills repository. It picks a method per task: new documents with the docx npm library, existing ones unzipped and edited in word/document.xml, and content read by converting to Markdown with pandoc. Scripts handle accepting tracked changes, adding comments and merging text runs. Much of the guidance is a list of docx-js pitfalls, such as A4 being the default page size, tables needing widths on both the table and every cell, bullets coming from numbering configuration rather than typed characters, and tables of contents only picking up built-in heading styles. After writing a file, the agent renders it to PDF with LibreOffice and to page images to check the result.

Requirements: Node.js with the docx package, plus Python, pandoc and LibreOffice for reading, conversion and rendering.

Trade-off: the licence is proprietary, so read the bundled terms before reusing the files. The toolchain spans several runtimes. The directory's automated check passed.

Skills built on the same docx-js approach

eigent-ai/docx

Best for: the official skill's jobs, plus a clean copy of a document with every tracked change accepted.

eigent-ai/docx reads through pandoc or unpacked XML, writes new documents with docx-js and validates them with a script, and edits through an unpack, edit and repack cycle. It also converts legacy .doc files, renders a document to page images, and accepts all tracked changes through a script that needs LibreOffice. It recommends Arial as the default font and warns that docx-js defaults to A4.

Requirements: Node.js with docx, Python for the scripts, pandoc, and LibreOffice for conversion and accepting changes.

Trade-off: its licence is listed as proprietary and the included LICENSE.txt governs reuse. The approach closely follows the official skill, so choose between them on licence terms and source. The automated check passed.

thinkinaixyz/docx

Best for: editing someone else's document with tracked changes.

thinkinaixyz/docx opens with a decision tree. Reading uses pandoc with tracked changes shown, or raw XML when you need comments, structure or metadata. New documents use docx-js, and the agent must read the whole docx-js guide before writing code. For your own document it uses basic OOXML editing, and for someone else's it recommends the redlining workflow by default, which the skill requires for legal, academic, business or government documents.

Requirements: pandoc, Python to run the unpack script, and Node.js with docx for new documents.

Trade-off: the licence is proprietary with a bundled LICENSE.txt. The directory's automated check reported informational findings only, the mildest level, so open the skill page to see what was noted.

Skills built on python-docx

hkuds/docx

Best for: Python environments where the agent runs complete scripts and python-docx is already installed.

hkuds/docx is licensed Apache-2.0 and works in two tiers. The default uses python-docx for creating, reading and simple edits. The advanced tier edits raw OOXML for what python-docx cannot express, which are tracked changes, comments and edits that must preserve every untouched byte. It notes that doc.paragraphs skips text in tables, headers, footers and text boxes, and that python-docx ignores tracked changes, so those must be read from the XML. It writes a new output file unless you explicitly ask for a replacement.

Requirements: Python with python-docx.

Trade-off: it is not meant for PDF, spreadsheets, slides or Google Docs. The automated check passed.

tokenrhythm/docx

Best for: filling placeholders and changing text while keeping the document's existing styles.

tokenrhythm/docx is licensed Apache-2.0. The agent chooses up front between inspecting, editing in place or creating. An inspect script dumps paragraphs, tables and styles as stable JSON that can be compared before and after an edit. An edit script applies operations such as replace_run and replace_text at the run level, so fonts and theme settings survive. Page layout, numbering and tracked changes are handled by patching the XML directly.

Requirements: Python with python-docx.

Trade-off: it relies on bundled scripts, so read them before approving them to run on important files. The automated check passed.

dataelement/bisheng-docx

Best for: formal Chinese documents, such as official notices laid out to the GB/T 9704-2012 standard.

dataelement/bisheng-docx is licensed Apache-2.0 and written in Chinese for BiSheng's code executor. It handles Chinese fonts, tables of contents, page numbers and repeating table headers on A4 paper, and it decides between its route and the platform's own Markdown export tool.

Requirements: BiSheng's code executor enabled, with python-docx, which the environment already provides.

Trade-off: it is tied to one platform's sandbox rules, including no network and no pip install, so it does not suit other setups. The automated check passed.

Skills that treat Word as one of several formats

• alchaincyf/huashu-md-html (MIT) treats Markdown as the source. It converts PDF, Word, PowerPoint and Excel files to Markdown, then produces HTML, Word, PDF and EPUB, with python-docx for manuscripts sent to publishers. It needs Python and pandoc, with Playwright for PDF, and its instructions are in Chinese. It is a publishing pipeline, not an editor for an existing document.
• genspark-ai/genoffice (Apache-2.0) builds Word documents from Markdown or HTML, converts between formats and applies structured edits through the genoffice command-line tool. It needs GenOffice installed.
• sylinko/officecli creates, reads, proofreads and edits Word, Excel and PowerPoint files through the host app's OfficeCLI tool. The directory lists its licence as a licence file that does not match a standard licence, so read it on GitHub before adopting it.
• cherryhq/office-transform (AGPL-3.0) extracts Word paragraphs and replaces text as a copy and never modifies the source file. AGPL-3.0 has sharing obligations for networked use.

The directory also lists wecomteam/wecomcli-doc, which is for WeCom online documents. It can import a .docx file as an online doc, but it manages documents in that service rather than editing Word files.

All of these passed the directory's automated check.

How to install a Word skill

Copy the skill folder to the location your agent scans, or use its plugin or CLI route. The steps differ by agent: see Claude Code and Codex, and the Claude Code install guide for personal versus project scope. Confirm that Node.js, Python, pandoc or LibreOffice are present, depending on the skill.

Picking one

Use the official skill for general work if its terms suit your use. Use thinkinaixyz/docx when the job is redlining another party's document. Use hkuds/docx or tokenrhythm/docx for Apache-2.0 terms in a Python-only setup. If you are weighing document formats together, compare the guides to Excel skills, PDF skills and PowerPoint skills, or see the top skills list.