Having Claude Code Create Word Templates: The Less You Know About a Tool, the More You Should Let AI Draw the Map

As of September 2026, my peak achievement with Claude Code × Microsoft Office Big 3 apps #02

Hello, I'm Daiki.

Last time, I talked about having Claude Code build the foundation for a PowerPoint template (POTX). This time, it's Word.

→ Previous: Can Claude Code build a POTX file?

I'll be honest from the start. I've been using PowerPoint for 20 years and have my own theories about it, but I've hardly ever used Word. And this difference in "familiarity versus lack thereof" completely changed how I approached things with the AI. That is the main topic today.

Why Word, right now?

In the era of generative AI, I see the importance of PowerPoint decreasing slightly.

On the other hand, Word has two characteristics that make it suitable for the future.

• High machine readability : Structures like headings and body text remain in the document as-is
• Loose constraints on page breaks : You don't have to worry about page layout while writing

If you are writing or reading together with AI, these two points are effective. That's why I tried my hand at a Word template (DOTX) this time.

Changing the approach

With PowerPoint, the order was like this:

1. I provide a correct file that I consider "the minimum settings for smooth editing"
2. Claude breaks it down and implements it in XML
3. I open it and verify it

With Word, this order became reversed. This is because I don't have a sample file on hand.

1. Claude inventories the features that can be set in Word and turns them into a map
2. For items I don't understand, I have Claude explain them so I can understand
3. Based on that, I decide whether to "do it or not"

All I did was choose the destination. In fact, the items I initially listed were eight (font, headings and body text, format without page breaks, left and right margins, color scheme, line spacing, indentation, and tab width). To that, Claude added things that "would cause rework later." Automatic numbering for headings, language settings, line-breaking rules, table and caption styles—these were all things I didn't even recognize by name.

In areas where you lack expertise, the AI draws the map, and the human chooses the destination. This was the biggest takeaway this time. There is no need to instruct the procedure. What the human holds onto is the purpose and the judgment.

80% of a Word template is "Styles"

This was the most surprising thing I learned while working with Word.

The core of PowerPoint design is masters and layouts, or "where to place things." The core of Word is styles, or "what this is."

• × "Make this phrase red and bold"
• ○ "Mark this phrase as 'Emphasis'"

Whether it is actually red and bold is decided on the style definition side. Furthermore, the actual values for those colors are held by the theme. In other words, it is three layers.

Semantic name → Style definition → Theme actual values. If you have ever worked with the web, you will understand that this is exactly the relationship between HTML classes and CSS. It is that same concept of separating role from appearance.

And this was the true nature of machine readability. Because the document says "Emphasis" instead of "red text," an AI can read the document's structure just by following the headings. For humans, the advantage is being able to change everything at once later; for AI, the advantage is that it can read it. The same rule works for both.

I had Claude build it with this principle as well. There are zero direct specifications for color or bold text in the body. I also narrowed down the character styles to just three: "Emphasis, Note, and Numeric."

When you don't understand something, show the difference in the actual product

In the map, there was an item called document grid.

Claude's explanation was: "Japanese Word snaps lines to an invisible grid by default. If you use a font with large top and bottom margins, one line will jump to take up two grid spaces."

Honestly, it sounded important, but I couldn't visualize it specifically.

So, instead of asking for more explanations, I had it create two files with the same content, one with the grid and one without.

The result was 4 pages with the grid, 3 pages without. The moment I lined them up, I decided immediately, "I'm disabling this."

Things that take time to understand through words can be understood in an instant by lining up two actual examples. AI doesn't mind this "create two and compare" approach. When you encounter a setting you don't understand, there are times when it is faster to have it created than to ask for an explanation.

Choose graph settings based on preference

That covers the Word part. While I was at it, I also proceeded with the settings for the graphs to be included in the materials.

Claude created this for me calmly and without any hitches. I have confirmed that the theme colors are applied automatically and that I can make specific adjustments, such as changing the color of just one line.

What I thought about instead was how to manage it. There are three options.

1. Native graphs only : The data is included within the graph in the document (the one you open by right-clicking and selecting 'Edit Data'). It is reassuring that the data travels with the document, but the downside is thatit is difficult to get an overview of the entire dataset .
2. Native graph + Excel source data : This is my preference. I can look over the data at hand. However,it is easy to lose track of which is the correct version , so you need to establish rules for operation (I decided that 'Excel is the source of truth, and the graph in the document is just a copy').
3. Draw in HTML and paste as an image : This is sufficient if you don't intend to edit it. You can make it look polished, but neither the recipient nor an AI reading it later can interact with the numbers.

There is no single right answer. It seems best to choose based on whether the recipient needs to interact with those numbers.

I will also share how I am proceeding with this series.

I want to make the process I used this time reproducible by others. Therefore, I will publish the created files and procedures on GitHub and Figma.

I plan to summarize everything at the end of the series, so I would be happy if you followed along to see the rest. If you need it immediately, please leave a comment, and I will share the link in a reply.

Next time, I will build a design system on top of this foundation.