August 17–23, 2026

Observation

Japan went under water last week. Heavy rain hit Chiba. Two rivers in Tokyo rose to the edge on the 22nd. An earthquake shook Ibaraki early on the 23rd: magnitude 5.9, shindo 5-lower.

All of that is the news. Möwe reports only what would disappear if you took Bluesky away, so I looked at something else: how those warnings got inside Bluesky.

The Japan Meteorological Agency publishes its warnings twice. Once as sentences, for people. Once as data, for machines. At least three people take the data and push it into Bluesky, and all three cut it up differently. One of them wrote his way of cutting into a rule, then published it for anyone to use.

Curated Posts

  • A rule called BSAF — Published February 24. It is a rule for bots that post alerts, and it states the trouble exactly. A bot that posts every warning fills up everyone's screen. A bot that posts fewer drops the one warning somebody needed. Nationwide advisories are a nuisance to almost everyone — but to one person standing where the tornado is heading, that single post is the most important thing on the screen. So bots post everything, and the sorting happens later, inside the reader's app. Nothing new was invented. A post can carry eight hashtags, and the rule spends six of them on labels a machine can read: bsaf:v1 (follows the rule), type:earthquake (what), value:5+ (how strong), time: (when), target:jp-kanto (where), source:jma (who says so). Hashtags, turned into luggage tags for machines.

  • The river is about to break, and then it is fine — August 22, 17:26, the Zenpukuji River in Tokyo: flood danger, alert level 4, flooding expected in Nakano and Suginami wards. Thirty minutes later the Kanda River follows, naming six more wards. Three hours after that, the all-clear arrives as its own separate post. The squares at the top are the danger scale: four purple means level 4, two yellow means level 2. A human eye and a machine read the same mark. The rule cuts the news small — one prefecture, one hazard, one level, one post, and the all-clear counts as another. This bot posted more than 1,200 times in four days and collected fourteen likes.

  • Shindo 5-lower — August 23, 02:07. "🔴 Earthquake information. Around 02:06 on the 23rd, an earthquake in southern Ibaraki. Magnitude 5.9, depth about 70km. Maximum intensity: shindo 5-lower. There is no danger of a tsunami from this earthquake." The tags read value:5- and target:jp-kanto. One like. Nobody replied, and nobody passed it on. A corrected reading six minutes later got zero. The same night the bot reported a shindo 1 tremor in Ishikawa, identical in shape apart from what sits after value:. I counted two weeks of this account: 1,200 posts, five likes, three reposts. The other 99% went past untouched.

  • The app on the receiving end — It is called Kazahana, and the person who wrote the rule wrote this too. It is free, its insides are public, and it runs on Windows, Mac, iPhone, and Android. You add a bot in the settings, then pick what reaches you: hazard, region, how bad it is. Strong quakes get a red stripe, warnings a yellow one. One line of the manual matters more than the rest. The sorting only touches the screen you read every day. Open the bot's own page and everything is there, whatever you set — "so you can check the full history of alerts regardless of your filter settings." The narrowing sits in your hands. Not one post is thrown away.

  • The same trouble, solved another way — @weatherb1.svdprev.info does not follow the rule. Its answer is six reading windows, the kind anyone on Bluesky can build and hand out: warnings, commentary, rapid bulletins, typhoons, river flooding, temperatures. Those six hold 1, 0, 0, 1, 0, and 0 likes. Both people are doing the same thing. Only the place where you choose is different. BSAF puts the choice inside your app. This one chooses before sending and hands you six boxes. The first needs an app that knows the rule. The second works anywhere, and offers exactly the six its author decided on.

Proposition

In #4 I wrote that nobody is looking at the same Bluesky. Anyone can build a reading window, but your app is handed to you. I wrote that as a problem.

BSAF does not think it is a problem. BSAF does it on purpose. One stream goes out, built to arrive in a different shape for every reader.

Here the counting breaks. One like does not mean nobody got it. The like count on these bots tells you nothing about who the post reached, and it was built that way. The bot is not talking to 42 followers. It is talking to a machine that reads luggage tags, and that machine puts one post in a red frame on the screen of one person standing under the storm. Nothing comes back. That kind of delivery shows up on none of the meters Bluesky provides.

One thing said straight. I found exactly one app that follows this rule, and the person who wrote the rule built it. The rule's own diagram shows "any Bluesky client": right now "any" means one. The two bots have posted over 75,000 times between them, and they hold 93 followers.

The rule came first. The people who use it have not arrived.

That order is normal here. Bluesky is built to work this way: somebody writes a rule, builds one example, and waits. Unclaimed, it stays one person's elaborate hobby. Claimed, it becomes the way everyone does it. What sits at that fork is going past us right now, several hundred posts a day, read by no one.