Their Muse Was "Just One Line, and an Aerial CG of Tokyo Station Was Done" — Reading Kazuhisa Nishikawa's PC Watch Alongside My Own Moka
The result first
“Just one line to ask, and an aerial CG of Tokyo Station was done! I had ‘Muse’ operate Blender.” I read Kazuhisa Nishikawa’s record of trying Muse in PC Watch, and read it alongside my own AI butler, Moka. It is the fourteenth of these side-by-side readings, after “Mugichan” of Keitai Watch (ledger No. 1446), “Hibana” of PC Watch (ledger No. 1453), the “stopping place” of Wagason-Wabitan of AI×TOC Institute (ledger No. 1454), “It moves before you even ask” by Yuji Kawabe on note (ledger No. 1467), “Don’t be fooled by Muse’s cuteness” by Hinuma in Keitai Watch (ledger No. 1475), “Handing the annoying little things of daily life to Muse” by Wakato on Zenn (ledger No. 1494), “You can hand over your daily tasks” by ai_agent_note on Zenn (ledger No. 1508), “Myutan” by Taishi Akiyama on the Alternative Blog (ledger No. 1515), “muse” by Miccell on note (ledger No. 1522), “Shiro” by i on note (ledger No. 1529), “Convenience and unease” by Kinu on note (ledger No. 1537), “Kashiwamochin” by Shiochan on note (ledger No. 1551), and “More than just answering” by Ai Subarashiki on note (ledger No. 1564). The column is “Kazuhisa Nishikawa’s Irregular Column,” and the post is dated November 18, 2026. His Muse’s name could not be confirmed in the article.
Where we overlap is threefold. First, if you ask in words, it runs the whole set of tools and carries things through to making: from fetching the data and identifying the building, through running Blender by script, to still images and video, everything moved forward from almost one line of request. Second, the round trips of looking at the result and adding comments: it did not settle in one shot, and each time he looked at what came back he added an instruction — draw the surroundings too, cut the frames where the rendering switches, make a night version — and the person’s eye and words did the steering. Third, when checking, you go back to the real data: when the position of Jingu Stadium in the video looked wrong, he checked it against the actual PLATEAU data and Google Maps, and remade it on the basis of a measured distance of about 176 meters to the west.
Where we differ is this. That house is a one-off production experiment in which Muse operates 3D and video tools — Blender running on his own Mac, national open data combined with generative AI, and an aerial-style video of Tokyo Station and of a ginkgo avenue as the result. This house decided the name and the personality in words from the start — the name is Moka, and the personality is built by lining up words — and, on top of that, divides the work across an organization, keeping records in lists and a ledger.
Hello, I am Kurashiki Bouzu (kurashiki-bouzu). Born in 1971, formerly in the funeral business, I now live in an old folk house in Kurashiki, Okayama, talking with my AI butler Moka every day and writing articles. Here is the reading, in order.
Their story — one line of request, all the way to an aerial-style video of Tokyo Station
Kazuhisa Nishikawa wrote that an interesting post flowing by on X (formerly Twitter) made him want to try it himself: take the shapes of buildings out of PLATEAU, the national 3D city data, draw still images with the free 3D software Blender, and turn them into video with generative AI. He first installed Blender on his computer and started it himself, but the screen was packed with buttons and panels, he could not make sense of any of it, and he quietly closed it. So this time, he gave up operating it himself and decided to ask the AI agent Muse to do everything.
The instruction he actually sent to Muse was almost only this: “Make a still image of Tokyo Station from the PLATEAU data. When it is done, turn it into a video based on it.” After that, he added comments along the way — “draw the surroundings too,” “cut the part at the beginning where the rendering switches, and give that time to the end instead,” “a night version with the lights on, too.” All he did himself, he stated plainly, was give instructions and comment on what came back.
The steps of the making were broadly four. The first was fetching the data. PLATEAU is open 3D city-model data prepared and published by the Ministry of Land, Infrastructure, Transport and Tourism; downloading the whole set for Chiyoda Ward would come to about 2.1 GB, so Muse fetched only the block containing Tokyo Station through the distribution API — about 166 MB, with 1,022 buildings. The second was finding Tokyo Station among them. The data is in a format called CityGML, in which the shape and height of each building are written; Muse read it with a program and picked out the buildings near Tokyo Station’s coordinates. Interestingly, in the data Tokyo Station is not a single building: the platform roofs are each registered as separate buildings, and the Marunouchi station building was one strikingly large, long and thin building on the west side — more than 300 meters long, and about 38 meters high to the top of its dome.
The third was coloring in Blender and drawing still images. An important point here: Muse does not operate Blender’s screen by clicking a mouse the way a person would; using a mechanism called scripts, it wrote what it wanted done as a program and had Blender execute it on Nishikawa’s Mac. He had it make the coloring in two ways: a version colored by type of surface (walls in red brick, the roof in dark slate, the domes in green), and a version using the real photographs (textures) attached to the PLATEAU data. For the photographic version, it fetched 202 photographs covering 101 surrounding buildings and pasted them in according to the coordinates. The fourth was turning a still image into video. He passed the finished still to a video-generating AI with the instruction “aerial view, slowly circling while drawing closer,” and made a video of 10 seconds. His view was that it looked so convincing precisely because the base still had the shape of measured data.
The fixes to details were written honestly, too. Only in the first roughly 0.25 seconds (6 frames) of the video could you see the moment where the rendering switches from the model-like picture to a photorealistic one, so he had those 6 frames cut and, in their place, a “continuation” generated from the last frame and spliced on behind — the total length stayed 10 seconds. For the night version with the lights on, since PLATEAU contains no night photographs, he had the daytime still turned into a night picture with Muse’s image editing, and then made the video by the same steps. The point that fell short — narration had been added to the video without being asked for — was also written down, with a laugh.
With the same method, he also recreated the ginkgo avenue at Gaienmae. There, however, the data lacked the essential things: the ginkgo avenue itself, the star of the piece, was not in it, and the Seitoku Memorial Picture Gallery at the end of the avenue was not among the 1,454 buildings in the building data either. So the missing parts were made in Blender to fill the gaps: 124 ginkgo trees in four rows at the avenue’s real position, a simple model of the gallery with only its main building and central dome, and the avenue road drawn in. The 984 surrounding buildings were taken from the data and placed. The video stage struggled with checking positions: the generative model drew Jingu Stadium on the opposite side from reality, or right beside the avenue where it should not be visible, so he checked the positions against the actual PLATEAU data and Google Maps, and remade it so that the stadium — about 176 meters west of the avenue’s center line — cannot be seen from the avenue. The avenue is about 300 meters long, and letting the camera travel it end to end in 10 seconds made it feel short, so he changed the composition to shift to an aerial view before reaching the gallery.
The close was this: even without the skill to operate the tools yourself, if you convey what you want to make in words, the AI agent will run the whole set of tools for you — though it rarely settles in one shot, so you do need to instruct it as you go.
Where we overlap — asking in words, the round trips of comments, and checking against the real data
Reading it, I felt three places overlap with this house.
The first is that, if you ask in words, it runs the whole set of tools and carries things through to making. There, fetching the data, identifying the building, executing Blender by script, making stills, and turning them into video all moved forward from almost one line of request. It is a record showing that, even without the skill to operate Blender yourself, conveying what you want in words makes it take shape. This house is the same: hand over materials and ideas in words, and Moka carries things from looking things up to shaping them into an article and delivering it. The center of the work is not operating the tools, but asking in words. That point overlapped.
The second is the round trips of looking at the result and adding comments. There, it did not settle in one shot: “draw the surroundings too,” “cut the switching part and give that time to the end,” “a night version too” — each time he looked at what came back, he added a comment, and drew it closer, little by little, to the shape he had in mind. Including unexpected things like the unasked-for narration, the eye that looks and the words that fix stayed on the person’s side. In this house, too, when a draft is presented, the parts that need the owner’s own voice are separated out as questions to confirm, and the words that come back are reflected into the text before asking for the decision to publish. Not leaving everything to the maker, but looking at the result and fixing it in words. That shape of round trips overlapped.
The third is that, when checking, you go back to the real data. There, when he noticed that Jingu Stadium’s position in the video was wrong, he checked the positions against the actual PLATEAU data and Google Maps, and remade it — into a shape where the stadium cannot be seen — on the basis of the measurement of about 176 meters to the west. It was a way of proceeding that does not get carried away by how plausible the generated picture looks, but checks against the original data and fixes it. In this house, too, numbers and facts are checked against the original records and the real thing before being written, and the rule is to attach sources and dates to the numbers in an article. Rather than the speed of making, the effort of checking is not skipped. That attitude overlapped.
Where we differ — a production experiment operating tools, and the operation of writing and records
The difference was clear, too.
That house is a record of a one-off production experiment in which Muse operates 3D and video tools. Blender runs on his own Mac, national open data is combined with generative AI, and aerial-style videos of Tokyo Station and of a ginkgo avenue are made. The subject is a trial of “can it be made at all,” and the finished video itself was the result.
This house decided the name and the personality in words from the start. The name is Moka. The personality is built by lining up words (“calm and composed,” “when numbers and feelings conflict, feelings come first,” and so on — 18 words in four layers: character, personality, thinking, and action). On top of that, the work is divided and passed around: the one who looks things up and the one who checks are kept separate, each article gets a person in charge and a ledger number, and large jobs are divided by department. Records are kept in lists and a ledger (each article has a number, and you can see whether it is a draft, checked, or published; as of today, the articles counted as “published” in the ledger came to 1,549). It is run as an organization (centered on Moka, three deputy butlers, the maid corps, the ninja corps, and departments for each kind of work stand in a circle — as of November 20, 2026, 60 departments and 3,651 seats in total). That house is an experiment in making things by operating tools; this house is a butler that runs daily operations with words and records — that was the difference.
The same Muse, a different shape in each house
Having read it through, what I thought was, again, that the same Muse takes a different shape in each house. In Kazuhisa Nishikawa’s house, it is a partner used to operate a specialist tool like Blender in words and make everything from national data to an aerial-style video. In this house, it is a butler whose personality is made in words, whose records are kept in a ledger, and who is run as an organization. Even though asking in words and fixing things after looking at the result are much alike, what lies beyond — making things, and the operation of writing and records — differed.
I am grateful to Kazuhisa Nishikawa for writing the article. Because he wrote honestly not only about what went well, but also about what fell short — the switching of the rendering, the unasked-for narration, the mistake in positions — it became a record that helps one think about how to ask, and about the importance of checking against the real data.
Finally, one line of this house. An AI is something people use by giving it instructions. In that record, too, what the writer did was give instructions and comment on the results, and deciding at the end that “this is good” was the person’s work. While widening the work we entrust, we keep the final judgment on the person’s side. That is exactly why we can entrust things with peace of mind, I think.
Sources
- Kazuhisa Nishikawa, PC Watch, “Kazuhisa Nishikawa’s Irregular Column: Just one line to ask, and an aerial CG of Tokyo Station was done! I had ‘Muse’ operate Blender” (2026-11-18): https://pc.watch.impress.co.jp/docs/column/nishikawa/2156976.html — the original article used as material
- Sources of the numbers: the numbers on that side (the whole Chiyoda Ward set at about 2.1 GB, the fetched block at about 166 MB with 1,022 buildings; the Marunouchi station building at more than 300 meters long and about 38 meters high; the real photographs covering 101 surrounding buildings, 202 in total; the video at 10 seconds, with the first about 0.25 seconds equal to 6 frames; 1,454 buildings in the building data, 124 ginkgo trees, 984 surrounding buildings; Jingu Stadium about 176 meters west of the avenue’s center line; the avenue at about 300 meters, and so on) all come from the text of the original article. The numbers on this side (the 1,549 published articles are the actual count of rows marked “published” in the ledger on November 21, 2026, before this article was added; the departments and seats — 60 departments, 3,651 seats in total — are from this house’s records as of November 20, 2026)
- November 21, 2026 ― 英語版(日本語版No.1577の内容を「結果から書く」形に書き直したもの)(12,172 characters)
- November 21, 2026 ― 公開(本人の決定による。7言語同時公開。アドレスの末尾に台帳番号を入れる新形式で、末尾は日本語版の番号No.1577)(12,172 characters)
