Dividing the Work Between a CMS and Generative AI

Generative AI makes a first draft quickly. At the same time, reusing figures, venue names, roles, and attendance numbers from past articles spreads factual inconsistency across a whole site. Using the operation of the site for Miyako de IT, an IT engineer community in Kyoto, this article organizes how the work is divided between a CMS and generative AI. It is written for operators who keep publishing event announcements, event reports, operating policies, and sponsorship information.

Conclusion: AI up to the draft, people for the publishing decision, the CMS for changing numbers

Here is the conclusion about the division of work.

• Generative AI: shaping material and producing a first draft. We assign only work whose errors a person can catch before publication
• People: checking dates, venues, roles, attendance, permissions, quotations, and attribution. We do not automate this
• CMS: holding values that keep changing, such as activity statistics and operator information. We do not copy them into article bodies

The value of generative AI is not that publishing becomes automatic. It is that responsibility for drafting, checking, and publishing is separated, so the operating record can keep going.

Background: the Miyako de IT site does not run on WordPress

We disclose the primary information first. miyakodeit.com is built with Next.js and does not use WordPress. Article bodies are held as Markdown strings in a single data file inside the repository and converted to HTML at build time. There is no admin screen for posting.

That said, the way of dividing work described here applies just as well to WordPress and other general-purpose CMSs. The difference is where things are stored; who checks what does not change. This is not an article about whether one CMS or platform is better than another.

Work we leave to generative AI

What we assign is the work of shaping raw material, such as the following.

• Organizing voice memos (turning spoken notes taken right after an event into bullet points)
• Listing candidate headings
• A first draft of an event report
• Shortening text per social platform (adjusting length and tone)
• Finding candidate related articles among past posts
• Detecting inconsistent terminology (for example “engineer” versus “IT engineer”)
• Extracting update deadlines (taking stock of descriptions tied to a point in time, such as “2026 edition”)

What these have in common is that a person can catch an error before publication. Work where errors slip past human attention is separated out in the next section.

Items a person always checks

We never pass the following straight through from a generative-AI output. A person performs the check.

The distinction between roles needs particular care. Miyako de IT records the relationship category for each partner separately, such as co-organizer, venue provider, or educational collaboration. Generative AI sometimes rewrites “co-organizer” as “organizer” from surrounding context. Because rewording a role bears directly on a partner's credibility, a person always checks it.

Attribution of results works the same way. The team that won HACK+2023 in April 2023, organized by NEMTUS, the NEM Technology Promotion Association, an NPO corporation, was made up of members who had connections through Miyako de IT. The award was made by the organizers, so we do not treat it as a project born from the community; we record it as activity by people connected to it. How the team came together is left to the HACK+2023 record. Notes of this kind are the first thing to disappear when summarizing is delegated. The details are collected in the HACK+2023 record and the Kyoto IT Community White Paper 2026.

Keep changing numbers in one place on the CMS side

Miyako de IT gathers its activity statistics into a single constants file, and every page, every article, and the API reads from it. Writing figures directly into an article body is prohibited.

The operating rules are as follows.

• The primary source for the figures is public connpass data, and a weekly job updates the snapshot
• We hold “published events” and “completed events” separately. The published count includes cancelled and upcoming events, so sentences about our record use only the completed count
• Years of operation are calculated from anniversaries passed since the founding date, the first co-working session on February 24, 2019. We do not write a fixed value

The figures in this article come from those same constants. At present that is 7 years of operation, 153 completed events, and 623 registered connpass members, with collaborations spanning 14 organizations. The latest values and their breakdown are published on the activity statistics page.

Statistics are not the only information kept in one place on the CMS side.

• Activity statistics
• Operator information
• Contact details
• The code of conduct
• Venue information
• Sponsor information
• Article update dates

Copying these into article bodies means every article has to be corrected whenever they change. Anything missed simply remains as a factual inconsistency.

Updates stop for reasons other than being unable to write

This section reflects our own experience of operating. When updates to a community site stop, the cause lies less in the effort of writing than in the checking step depending on one particular person. If only one person can confirm dates and roles, publication stops during the weeks that person is busy.

Introducing generative AI makes first drafts faster, but it does not clear the bottleneck in checking. If anything, more drafts lengthen the queue waiting to be checked. We recommend deciding the division of work before introducing the tools.

Steps for trying this division of work (our operating guideline)

The following steps are a Miyako de IT operating guideline, not a universal answer. Adjust them to your own operating structure.

1. List the values in your articles that change (statistics, prices, venues, contact details) and move them into one place on the CMS side

2. Fix the items a person checks as a pre-publication checklist

3. Limit the work handed to generative AI to what is not on that checklist

4. Separate drafting from checking, and have at least two people who can check

5. After publication, set a date to review descriptions tied to a point in time, such as “2026 edition”

The measurement design is covered in how we measure a community site with GA4, and choosing an announcement platform in choosing between connpass, Doorkeeper, and Meetup.

Summary

• Generative AI shapes material and produces a first draft. People hold the publishing decision
• Dates, venues, roles, attendance, permissions, quotations, and attribution are checked by people rather than automated
• Values that change are kept in one place on the CMS side and not copied into article bodies

Co-working sessions and study sessions suit bringing operating questions to other people. Upcoming dates are on the event list, and newcomers may also want to read the first-time guidance.

Related articles

• How We Measure a Community Site with GA4
• Six Months of Turning a Member Site into a Publishing Site
• How to Start an IT Community on connpass