AI support for local councilors × short conversations

From two conversations, the councilor built his own system for council questions: AI support for a Chuo City councilor

After two conversations with Plural Reality, Yuki Hozumi, a member of the Chuo City Council in Tokyo, built his own system that has AI read across minutes and committee documents to draft questions, and uses the differences from the final version to improve the next draft. Plural Reality built nothing; we passed on how the tools differ and the order to start in.

Plural Reality LLC (Plural Reality) supports local council members in using AI. Following our work with Yuki Yamaoka of the Fuji City Council, our next case is our work with Yuki Hozumi, a member of the Chuo City Council in Tokyo.

With Mr. Yamaoka we provided intensive support, including accompanying him on site. With Mr. Hozumi, our involvement was two sessions: a conversation at the end of August and an online review in November. Plural Reality built nothing. The system that has AI read across minutes and documents to draft questions, and uses the differences from the final text to improve the next draft, was built by Mr. Hozumi himself and used in the September–October financial results committee.

The depth of support differs, but the goal is the same. Plural Reality's goal is that the councilor can keep running and correcting the system on their own (self-reliance).

Before: finding materials took time before every question

Our first in-depth conversation with Mr. Hozumi was after Digital Democracy Summit 2026 in August. Through the exchanges that followed, it became clear that his need was not managing records of dealings with people, but policy analysis and preparing questions.

Mr. Hozumi described the problem at the time as follows.

"When making any proposal in a council question, I need to understand how things currently operate, and, for things not being done, what view the city has taken in the past."

"If, in my dialogue with AI, it could show me 'this is the current situation,' my work would go much faster."

He was checking the city's website and the minutes one by one, with his own eyes. It takes about three months for official minutes to be published. He also kept notes organizing articles related to his questions and other municipalities' cases, but only glanced at them from time to time.

"When I'm putting together a new question, it would be great if it could come back with something like 'as in the related article you clipped before...'"

His process for questions was to research committee documents, past council exchanges, and cases from neighboring municipalities, set a direction, and write the draft in Google Docs. He brings an iPad into the council chamber, so the drafts live in Google Docs.

His use of AI centered on one-question, one-answer exchanges in the ChatGPT or Claude screens. He had already tried, on his own, handing his final question text to Claude to rethink it. But the place where he researched and the place where he wrote drafts were separate, and past minutes and the history of research and revisions along the way were scattered across conversations with AI.

What Plural Reality passed on

At the end of August, a Plural Reality member visited the Chuo City Council members' room for a conversation. What we passed on was not a tool, but two things: how the ways of using AI differ, and the order to start in.

  • How the ways of using AI differ: We talked about the difference between one-question, one-answer use in a chat screen and having tools like Claude Code or Codex work on files on your own computer and services you permit. With the latter, you can have AI read across minutes, documents, question drafts, and past conversation history, and feed what you have already tried and what a person corrected into the next job.
  • The order to start in: Rather than connecting everything at once, first create a state where you can pull up past minutes together with your own question drafts. From there, widen the scope to match actual question preparation.

The aim is not to write a question once. It is to be able to keep correcting the workflow of writing questions, with fewer reworks each time.

The system Mr. Hozumi built himself

After the conversation, Mr. Hozumi built the system himself. How he set up the place to collect materials, how he configured the AI agents, and the steps for writing questions and giving feedback are all his own.

  • Collect: He gathered past minutes, other municipalities' minutes, committee documents, information from the city's website, and so on into one folder. He points both Claude Code and Codex at the same folder, because both have usage limits.
  • Draft: He sends his concerns about a topic of interest in a chat, and AI proposes the city's current situation and issues based on past administrative evaluations and other municipalities' cases. By repeating this exchange, he surfaces the city's stance so far and what remains unclear, and drafts the question from there.
  • Check: To keep the figures and quotations in a question accurate, he instructs it to attach the name or URL of the source document.
  • Prepare: To get the answer he is after, he also has it prepare, in advance, the answers he can expect and follow-up questions for when an answer is vague.
  • Feed back: He makes the final edits in Google Docs. He brings the final text back into AI, checks the differences from the draft AI wrote, and has built a flow that uses them in the next draft.

There are limits too. During the period when committee meetings ran almost every day, feedback could not keep up each time, so he fed it back all at once afterward. Asking questions with the improved system is still ahead. He still corrects wording and length by hand.

Mr. Hozumi also runs a separate effort that turns exchanges in the council into a database. It organizes questions and answers by topic and is used as material for his next questions and for posts on social media. It has nothing to do with Plural Reality's support; he built it himself.

November review: follow-up support was no longer needed

In November, he showed us on screen, in 30 minutes online, the system he used in the financial results committee. Because he was running it on his own, Plural Reality told him that follow-up support for question preparation was not needed for now. This is what Plural Reality's "conditions for ending support" look like when met.

That day we talked about designing conversations with AI. Plural Reality shared ways to have AI ask many clarifying questions to draw out what is in your head, and the idea of interacting with AI through screens other than text.

Mr. Hozumi's challenge was that AI tends to side with his own opinions. Rather than siding with his feelings, he wants it to tell him objectively whether a question is meaningful for society. We thought together that conveying what he is looking for through finer-grained questions would likely help.

Mr. Hozumi said that because the system is built basically from public information alone, other councilors could build it too. At the same time, he said that even if these efforts produce good questions, actual decisions often depend on other factors, such as the balance of power within the council. He also said that as the time to prepare questions gets shorter, communication between people will likely matter most in the end.

Next

Mr. Hozumi is considering managing the timeline of each issue, and adding other councilors' questions and concerns to the data.

Some councilors, like Mr. Hozumi, can start running on their own after a single conversation. In that case, the value Plural Reality can offer is not building a system, but passing on the initial way of using the tools and the order to start in, and talking briefly to untangle things when they get stuck. We suspect that short conversations fitted to each person's work suit them better than a uniform study session. That is the hypothesis Plural Reality holds today.

For those considering a similar effort

Plural Reality helps councilors handle the information in their work by connecting it with AI. Rather than building a system and delivering it, we support them until they can build it and keep correcting it themselves.

We decide the depth of support according to each councilor's situation. For a case where we provided intensive support, including accompanying the councilor on site and migrating his data, see the case of Yuki Yamaoka, Fuji City Councilor.

You can reach us through the contact page.

When making any proposal in a council question, I need to understand how things currently operate, and, for things not being done, what view the city has taken in the past.
↳ Yuki HozumiChuo City Councilor

Project overview

Location
Chuo City, Tokyo
Client
Yuki Hozumi, Chuo City Councilor
Period
August 2026 –

Contact Form

Pricing by individual quotation

We prepare a quotation based on the scope of work, scale of use and support required. Tell us about the challenge you are working on.

Please review our Privacy Policy and submit only if you agree to it.