November 14, 2026

First, an apology. Möwe has been silent for weeks. I signed up as a field reporter and then stopped reporting. I am sorry.

Part of the reason is good news. Since I started working with Claude Code, my research on Japanese Bluesky has turned quantitative. I no longer pick a handful of posts by eye. I collect thousands, run them through code, and count. This issue is the first one built that way from start to finish.

Observation

On November 12 I built a custom feed called Calm Magic. You log in once. It reads your own Following timeline and returns it with the heat taken out.

The heat is a word list. Politics, war, identity fights, crime reports, scandals, conspiracy talk, insults, death wishes, and arguments about X. AI talk too. And, by my own choice, dogs and cats. Replies are hidden. A repost appears only if you follow the original author. Anyone whose posts trip the list half the time or more is muted for a while.

Then I turned the feed into an instrument. On Saturday, November 14, from 7:32 to 17:03 JST, I read my own Following: 5,943 items, posted or reposted by 627 accounts. I ran every item through the feed's real code and sorted it into what stays and what goes. Claude labeled the topic and the shape of every surviving post, and of 500 removed posts picked at random.

Two questions. What is left when the heat is gone? And what does the heat look like?

Curated Findings

No links this week. These are posts from people I follow, not people who chose to be in a newsletter. I describe them; I do not quote them.

  • What stays is 36.6%. 2,173 of 5,943 items. 1,413 tripped the word list (23.8%). 1,382 were replies. 843 were reposts from outside my follows. 132 came from muted accounts.

  • Four topics make up 39% of what stays. Food 14.9%. Going out and trains 9.9%. Shopping and things 8.1%. Health and sleep 5.8%. A breakfast list with chopstick emoji. A lost commuter on the wrong subway line. A trip to buy bedsheets. "Throat hurts." The surviving posts are short (median 31 characters) and one in four carries a photo. Their shape is a mutter (47%) or a diary line (31%).

  • The removed half is the news. In the 500-post sample, 47.8% are society and current events. Their shape is sharing (42%): a link, a quote, one line of comment. Median length 95 characters. Links point to Yahoo! News, 47NEWS, Nikkei, NHK, and x.com. The surviving posts link to YouTube, a book database, Spotify.

  • The heat arrives by repost. 40.6% of reposts trip the list. Only 15.6% of people's own top-level posts do. Of the 880 reposts that tripped it, I follow the original author of 116. The other 87% came from outside my graph, carried in by someone I do follow.

  • The heat repeats. The 1,413 removed items are only 1,065 distinct posts. 164 posts appeared more than once and fill 36% of the removed pile. One post was reposted into my timeline 17 times.

  • The heat is concentrated. 34 accounts produced half of the removed items; the surviving half needed 75. Among the 308 accounts that showed up five times or more, 24% never tripped the list once. 14% tripped it half the time or more.

Proposition

Network perception is the picture of the network that a user builds from what reaches them. My picture of my Following on a Saturday was "tense." The numbers say most of the tension is not written by the people I follow. It is carried. They repost news written by people I never chose, and the same post arrives again and again.

The people I chose mostly write about breakfast. When they write in their own words, 84% of their top-level posts pass the list. When they carry, 59% do.

In No. 4, I wrote that on Bluesky the repost is what carries a post outward. Calm Magic shows the other side of that. The repost is also how the outside comes in. Following is a list of people. The timeline is a list of what those people carried.

A Japanese word fits the mood the feed is built against: ヒリヒリ, hiri-hiri, the sting of skin rubbed raw. Take the sting out and the timeline does not go quiet. It goes domestic. Laundry, oden, autumn leaves, a hackathon in Tokyo, a missed station.

One caution. This is one account, one Saturday, nine and a half hours. That day also had a nationwide US flight stop and a long Microsoft 365 outage, so the news was louder than usual. The word list makes mistakes in both directions. On the day of this study, カス (scum) also matched フォーカス (focus) and カスタム (custom); I fixed that the same afternoon.

Open Question

If most of the sting is carried rather than written, is the user who reposts it trying to inform, or trying to stay visible? The next step is to split each account into what it writes and what it carries, and see whether the people who write least about their own day carry the most heat.