"Hey, can't we just use AI to quickly auto-generate our instruction manuals?"

Lately, we’ve been getting this exact question a lot from DX managers and development teams. Looking at the rapid evolution of generative AI like ChatGPT, it certainly feels like we're living in an era where anyone can easily create text.

However, as professionals in the highly specialized manual production industry, we know the reality isn't quite that simple. While AI is fantastic at writing beautiful Japanese based on provided information, grasping the entire product structure from disjointed specifications and building a table of contents that won't confuse the end-user is incredibly difficult.

Traditionally, technical writers have relied heavily on years of "tacit knowledge" to handle this structuring process. They read between the lines of jargon-heavy specs and translate them into a context the user actually understands. Because this skill is practically invisible and difficult to pass down, the workload inevitably bottlenecks with a few veteran writers, making the process highly dependent on specific individuals.

For instance, imagine receiving this dry, robotic string of text from the development team:

• "Hold Time > 5s on power button transitions to pairing standby state."

• "UI feedback: LED indicator alternates blue and red."

While functionally correct, this cannot be used as a manual as-is. We technical writers first imagine a user trying to connect their earbuds to their smartphone for the very first time. We then translate that raw data into a living, actionable instruction: "Press and hold the power button for 5 seconds until the lamp flashes blue and red alternately." This process of "reading between the lines" is exactly the tacit knowledge we mentioned, and it's the biggest cause of workflow bottlenecks.

"How can we integrate this invisible tacit knowledge and structuring skill into AI to eliminate dependency on individual writers?"

That very question has been the main theme of our 3-year joint research project with Hokkaido University’s Harmonious Systems Engineering Lab (led by an amazing team including Prof. Kawamura, Specially Appointed Prof. Yamashita, and Assist. Prof. Yokoyama)（http://harmo-lab.jp/）. In this column, I am thrilled to introduce our concrete approach based on the contents of our recent reporting session.

Year 1: Text Optimization and the Next Big Wall

When we kicked off the joint research in 2023, our first year was primarily focused on "text creation." We programmed the AI with technical writing rules—such as "one sentence, one meaning" and "eliminating double negatives"—to verify its ability to automatically correct existing drafts into highly readable text.

Through this initiative, we gained solid confidence in "text optimization" for the Japanese language. However, when we tried to adapt this to the actual production floor, we hit a much higher wall: "grasping the product structure from zero to build a table of contents" and "filling in the contextual gaps of fragmented specs".

The "AI Haiku" Epiphany: Unpacking Context

Regarding the difficult task of "reading between the lines (complementing context)," Assist. Prof. Soichiro Yokoyama shared some highly fascinating research during the session on using AI to decode Haikus.

You might be wondering, "Why did YAMAGATA, a manual production company, partner with a Hokkaido University lab that studies Haikus?"

Actually, three years ago, when we first approached Hokkaido University, our reasoning lay exactly there. We hypothesized that our job of "delivering engineer-to-engineer context to the end-user" and "appreciating a Haiku" shared the exact same structure in terms of information transmission.

• Haiku: A poet compresses their emotions and scenery into a hyper-condensed 17 syllables. It’s packed with unspoken context, like the customs and feelings of the time, which the reader then "unpacks" using their imagination to appreciate.

• Manual Production: Developers compress massive amounts of technical data into "spec sheets." These are loaded with unspoken understandings (context) between engineers, often containing unnecessary information for the general user or presenting necessary information in impenetrable ways. A technical writer translates this, "unpacking" it into actionable instructions so the end-user can act without hesitation.

The system built by Prof. Yokoyama’s team does exactly this: it has the AI decompress and verbalize the unspoken context compressed into 17 syllables, auto-generating stunningly vivid descriptive explanations. In an actual collaboration with the Ehime Museum of Art（https://www.ehime-art.jp/lets）, visitors wrote abstract Haikus inspired by artworks, and the AI beautifully articulated the emotions and scenery behind them to serve as exhibition commentary.

Of course, mapping software specs to Haikus is just our own perspective. But to see that if AI can accurately verbalize the mood and background of a Haiku for a general audience, it should also be able to translate dry technical strings into end-user context—this was a session that firmly validated the hypothesis we held at the start of our joint research. I felt a great sense of relief and excitement seeing it take shape as a reliable technology applicable to practical business.

The 4 Steps to Auto-Generating Manuals from Specs

The person who spent three years carrying out the practical work of this research was Ryosuke Uemae, a 2nd-year master's student. He shared his latest results right after presenting them at the ICICT 2026 international conference in Hawaii（https://icict.org/）. He shared a rather grueling anecdote about being warned by the consulate to stay indoors due to consecutive days of rainstorms, which elicited a few sympathetic smiles through the screen, but his research directly addressed our most pressing practical headaches.

Mr. Uemae built a system to auto-generate a user-centric manual from messy information sources by having the AI sequentially execute the following four steps:

1. Assess Excess/Deficiency: Analyzes the spec sheet, removes internal specifications (noise) unnecessary for the user, and simultaneously points out missing prerequisites.

2. Generate Topics: Following a strict "one topic, one theme" rule, it splits the info by function or task and drafts titles and body text.

3. Generate ToC: It logically hierarchizes and sequences the generated topics based on the user's cognitive process (Learn -> Prep -> Use -> Troubleshoot).

4. Integrate into a Manual: Merges the Table of Contents and the topics, outputting them as a final manual.

What is truly noteworthy is the evaluation of the manual structure generated by this system. When evaluated by 140 manual production company employees and 8 professional technical writers, it recorded scores of "4" or "5" on a 5-point scale across many items.

It was highly praised by professionals for aspects like "the title matches the content" and "the sequence of the table of contents is logical." While not perfectly flawless, the significance of proving with data that AI can replicate a large portion of the "tacit knowledge of information design" that previously only existed in the minds of veteran writers is, I believe, absolutely massive.

The Future: Manuals as a "Knowledge Feedback Hub"

During the latter half of the reporting session, we demonstrated our prototype app, "Manual Generator," which implements these research results. When a product spec PDF was loaded, the AI instantly supplemented missing information (such as safety precautions) and auto-generated a table of contents tree and a set of topics based on writing rules.

Finally, Specially Appointed Prof. Tomohiro Yamashita from Hokkaido University’s Information Initiative Center spoke about the future vision for manuals.

Manuals have evolved from the print era (1st Gen), to digitalization (2nd Gen), to interactive AI (3rd Gen). After passing through the 4th Generation—where user query data is fed back to manual editors to improve the manuals themselves—Prof. Yamashita presented the future of the 5th Generation. This future expands the feedback loop beyond just manual editors, sending the data directly back to the product designers and developers.

If data on where users get confused is fed directly back into development, the UI will be improved in the next model, ultimately leading to the extreme conclusion of creating "products that do not need manuals".

While this is certainly a heavy theme to ponder for those of us whose livelihood is creating manuals, Prof. Yamashita's words resonated deeply: "Your jobs will not disappear; rather, your role will evolve from simply writing text to designing the entire architecture where knowledge feeds back". Future manual managers are entering a phase where they will evolve from mere document creators into "knowledge hubs" that connect products and users, updating the products themselves.

Start Small to Succeed with AI Integration

So, how should we utilize this AI-driven manual structuring technology in actual business settings?

Attempting to AI-ize the entire company's production workflow all at once is likely to cause confusion on the floor and raises the hurdle for gaining management approval. As we have frequently mentioned in our columns, the ironclad rule for establishing a new system and demonstrating clear cost-effectiveness is to "Start Small".

First, we recommend narrowing your target to something where the effects will easily show up in the numbers—such as a "highly revised product" or a "single volume that requires extensive multi-language deployment"—and test-introducing AI support for structuring and drafting the table of contents. Just having the AI handle the foundational work of a task that previously required building a structure from scratch by individuals will drastically reduce production lead times and the burden on the person in charge.

At YAMAGATA, we combine our long-standing knowledge of document production with the methods for utilizing AI cultivated through joint research and various document production projects, to propose DX (Digital Transformation) solutions that are highly practical and perfectly fit your frontline needs.

"Can AI really structure our company's highly unique specs effectively?"

"I’d like to try a Proof of Concept (PoC) in a small area first."

It’s perfectly fine if you don’t have a specific request ready. From organizing the issues in your current production workflow to creating a roadmap for internal proposals, please do not hesitate to consult with us.

contact us