The last clipboard in the smartphone era
Japan recently declared victory in its government "war on floppy disks." Another analog institution survives in its neighborhoods: the kairanban, a clipboard of printed notices that is passed from house to house, each household stamping or signing it before walking it to the next door. It can take a week or more to circle the block. Nobody knows who has actually read it. It is also, in its way, beloved — a small ritual of contact in an aging society.
Neighborhood associations (chōnaikai) — the volunteer bodies that run local festivals, disaster drills and community halls — are shrinking and disbanding across Japan. TV networks have spent the past year reporting the "problem" side: associations dissolving, autumn festivals losing their portable-shrine processions. Jag's Kairanban-app is an attempt at the "what now" side.
Who built it, and how
I am an active officer of my neighborhood association in Koto ward, Tokyo, and the representative of Jag Project, LLC. I have planned and directed software projects for most of my career in Japan's media industry — but the last time I personally wrote a program was BASIC games in junior high school.
I built Jag's Kairanban-app by directing a generative AI coding agent (Anthropic's Claude Code) in Japanese: describing what I wanted, deciding between options it proposed, and refusing to accept "done" without proof. A public beta took about two weeks. The app has been updated daily since, including contributions through GitHub. I developed it as a working proposal to my own association — which is still considering it — and released it as open source so any association can use it. We are now recruiting the first partner association to verify it in real use.
| Time to beta | About 2 weeks |
|---|---|
| Code | ~10,000 lines of TypeScript, 23 database migrations |
| Running cost | From ¥0/month (free-tier hosting) |
| Languages | Japanese, Easy Japanese, English, Chinese, Vietnamese |
| Development cost | Mainly paid subscriptions to AI tools (ChatGPT, Claude) |
The design premise is economic as much as technical. Electronic circular boards are becoming a market in Japan — a newspaper company has entered the business, and the Tokyo metropolitan government now subsidizes association digitization. Most offerings are monthly subscriptions, which means association fees flow out of the neighborhood every month. Jag's Kairanban-app inverts that: free infrastructure, open code, and if anyone is paid to maintain it, it can be the association's own IT-literate members or a local business. We call the idea "digital self-governance" — residents using digital tools to run their own community, rather than leaving digital infrastructure entirely to government or vendors.
The 20-out-of-100 manual
The hardest problem was not code. It was the button labeled "Add to Home Screen."
The app is a PWA — a web app that installs to the phone's home screen. On iOS, push notifications only work after that installation, and Apple provides no way to automate it: a member must tap through Share → Add to Home Screen → Add by themselves. So I had the AI produce a beautiful step-by-step manual with screenshots. I scored it 20 out of 100.
People who aren't comfortable with smartphones cannot memorize what they saw and then operate from memory.
On iOS, the moment you start the procedure, the system share sheet slides up and covers the instructions. A conventional manual therefore demands that the user memorize two or more steps — which is precisely what the least confident users cannot do. However polished, it fails by design.
The rebuild started from measurement. On real devices, we found the iOS share sheet
leaves the top 42.7% of the screen uncovered. We also confirmed that
JavaScript keeps running while the sheet is open, that a web page cannot detect the
sheet opening at all, and that navigator.share() cannot offer
"Add to Home Screen." The answer: a "zero-memory" guide that keeps photographs of every
button on-screen inside that 42.7% band, all steps visible at once, so the user only
compares and taps — nothing to remember, no timed auto-advance.
Even that failed once: the band was sized in fixed pixels, and on the smallest iPhone it silently stretched to 48% of the screen, hiding steps 4 and 5. Only a real-device screenshot caught it. The guide has since been extracted as its own open-source component, pwa-install-guide, with all the measurements documented.
Other failures, all published
- Two AI sessions, one repository. Running two AI chat sessions on the same codebase in parallel, one session committed the other's half-finished files. We now require each session to commit only files it changed itself.
- A test page left in production. A verification HTML file was forgotten in a public directory. Harmless content — but it became a rule, and then a script that checks mechanically before every release.
- Eyes are not an audit. Before publishing, we ran optical character recognition over all 33 images on the public pages and machine-matched the text against the database. Zero personal-data hits — but only the machine check could prove it.
- A race condition in production. The demo site's daily data reset ran twice concurrently and briefly wiped role data. It was found, reproduced under 15 parallel requests, and fixed with a database lock — while this document was being written.
What the AI could not do
Claude Code wrote essentially all of the code. It did not — could not — do the parts that turned code into something a community can use:
- Judgment. Deciding, for example, that login would use phone numbers only — trading security risk for the certainty that members who cannot manage passwords can still get in. That trade-off is documented publicly, with a standing request for better ideas.
- Verification on real devices. An AI reports "done." Whether step 5 is visible on the smallest iPhone is a fact about the physical world.
- Writing decisions down. AI sessions forget. Undocumented decisions get "helpfully" reverted by the next session.
- Knowing who it is for. The 20-point review existed because someone was watching the user, not the product.
Why publish the failures
We believe software use is shifting from "buy a finished package" to "take open source, have an AI adapt it to your own problem, and build it yourself." That future needs honest accounts of what directing an AI is actually like — not demos. This project is Jag Project, LLC's flagship verification of that future, and the record, including everything that went wrong, is the point.