Markdown to Word skill

No public download yet. The skill is written and its guidance has been tested, but it is not distributed anywhere you can fetch it from today. It is useless on its own — it drives either the command line tool or the MCP server, and both of those are held behind an unresolved licence question about the shared conversion engine, so none of the three has shipped. The instructions below describe how installation will work, not a directory you can copy right now. If you want the conversion itself today, the same engine runs in your browser on the Markdown to Word converter, with no install and no account.

What a skill adds that a prompt does not

A skill is a folder of instructions that a Claude client loads when the conversation looks relevant. This one carries three things a one-off prompt cannot: the trigger conditions that make Claude notice a document request in the first place, a route decision so it uses the best converter that is actually available, and a written list of capability boundaries so it refuses to claim a PDF it did not produce. The last one is the reason the skill exists in this form. An assistant that writes a .md file, renames it .docx and announces success is worse than one that says it cannot convert; the skill states that rule explicitly, along with the two dependencies that commonly fail.

How installation will work

A skill is installed by placing its directory in a skills folder — personal, so it is available in every project, or checked into a repository so a team shares it:

cp -r markdown-to-word ~/.claude/skills/     # personal, every project
cp -r markdown-to-word .claude/skills/       # project-scoped, committed

Then at least one runtime has to exist for the skill to drive. It will be either the command line tool or the MCP server:

npm i -g @mdtodocx/cli                                                  # CLI route
claude mcp add markdown-to-word -- npx -y markdown-to-word-mcp -w "$PWD"  # MCP route

Both commands depend on packages that are not on npm yet. Neither will succeed today; they are listed so the shape of the setup is clear, and so this page does not have to be rewritten when they land.

What you say to trigger it

The skill's description lists the phrasings that should wake it, deliberately including the ones that never mention Markdown, because that is how people actually ask. "Convert this to Word", "export as Word", "save this as a document", "give me a downloadable version of that report", "turn this answer into a document I can send", "make a .docx", "markdown to pdf", "give me a Word version", and in the other direction "read this .docx" or "turn this Word file into Markdown". It is scoped on purpose: slide decks and spreadsheets are explicitly out, so it does not fire on a request for a .pptx that it cannot serve.

How it picks a route

The skill checks in a fixed order. If tools named markdown_to_word, markdown_to_pdf, word_to_markdown or convert_file are already available, the MCP server is connected and Claude calls it directly: no shell, no install, and the workspace sandbox already applies. Otherwise it falls back to running the command line tool with Bash. If neither route exists, it writes the Markdown to a file, says plainly that it could not convert it here, and points at the hosted converter — it does not invent a file. Two more rules are written down because they are the ones an assistant gets wrong under pressure: never hand back a .docx that was not actually produced, and never rename a .md to .docx.

The workflow it follows

Content in the conversation is written to a real file first, because piping a long document through shell arguments is how quoting bugs are born. The conversion runs. The result is checked rather than assumed: the CLI prints one line per file, exits non-zero on failure, and names every degraded feature in its warnings, and the skill tells Claude to read those warnings and pass them on. The report back to you carries the output path, the size, and any warning that changed the document — a skipped image, an unrendered equation, an unrendered diagram. For structured documents it knows the flags that matter: a table of contents, numbered headings and page numbers for a spec, native equations for a paper, Mermaid diagrams for an architecture document, and a reference .docx to match a corporate template.

What it is told not to claim

Three boundaries are stated in the skill itself, because each one produces a confidently wrong answer if it is missing. PDF requires a headless Chromium; without playwright or puppeteer the command fails and Claude is instructed to convert to HTML and let you print it rather than assert a PDF exists. Mermaid requires the Mermaid CLI; without it, diagrams stay as code blocks and a note says so. And the free-versus-Pro split is a property of the hosted service described on the pricing page, not of a package on your machine — a locally installed CLI or MCP server runs the whole engine with no entitlement check, so the skill carries the tier table to advise you correctly instead of guessing. Remote images are a fourth, quieter case: in a sandbox with no network they are skipped with a warning, while local paths always resolve relative to the Markdown file.

Skill, MCP server or CLI?

They are three entry points to one engine, and the choice is about who is driving. The CLI is for you and your scripts: deterministic, pipeable, CI-friendly. The MCP server is for an agent that already knows it wants a document, and it adds the workspace path sandbox. The skill is the layer above both: it decides whether a document is what you meant, then uses whichever of the two is present. Installing the skill alone does nothing; installing a runtime alone works but means asking for the conversion explicitly every time. If you only want a document right now and none of this is set up, the browser converter is the shortest path, and the Chrome and Edge extension exports an AI answer straight from a chat page.

Related tools and guides