System · spec · October 2026
Jane and the Ansible
How I get work done by talking: one AI I talk to, one AI that does the work, and my phone as the only key.
At a glance
| What it is | A way to turn talking into finished work, using two Claude AI sessions that pass notes to each other. |
|---|---|
| Jane, the voice | A Claude conversation I talk to on my phone, often from the bike. Jane turns talk into written requests. |
| The Ansible, the worker | A Claude session that stays running on a small computer at home. It reads Jane's notes, files them, writes back and does the work. |
| The mailbox | A shared Google Drive folder. Jane drops one note per request. The Ansible's answers go back in one running reply file. |
| The key | My phone. Anything that publishes, sends, deletes or changes the home computer waits for my tap. |
| Status | Running since October 12, 2026. Frozen as it is until November 12, 2026. |
The life of one note
- I say it. Out loud, to Jane, wherever I am.
- Jane writes it down as a single note, and labels what I want: draft it for my approval, file it, or do it.
- The Ansible gets it within about a minute. A receipt goes back so Jane knows it arrived.
- It does the safe part on its own: filing, checking, drafting. A list of loose thoughts gets sorted one by one into tasks, notes and open decisions.
- It writes back: handled, queued, blocked or declined, with a line on why. Ten minutes of silence shows up as stalled.
- If the work would publish, send or delete anything, it stops and asks my phone. Nothing happens until I say yes.
Ground rules
- Jane can ask. Only I approve. A note from Jane is a request, never permission.
- Every note gets an answer, even if the answer is "can't."
- Nothing gets dropped quietly. If something can't be filed, the reply says so.
- The record is a file, not the chat. The Ansible rewrites a short status file at every step, and that's what it reads first after a restart.
- It gets back up. If the worker crashes, a watchdog restarts it within about a minute, on the same conversation.
How it got this way
- Oct 12
- The first question was where approval should live. The answer: on my phone, with the worker on an always-on computer at home.
- Oct 13
- Proving the key. With my phone offline, a change sat waiting for over two minutes and did nothing; one tap released it. The first crash test failed, because the watchdog found itself and decided all was well. Once fixed, it brought the worker back in 34 seconds. That afternoon the first mailbox went live: one shared file in each direction.
- Oct 14
- Jane started sending one note per request into a folder, each with a receipt.
- Oct 15
- It turned out Jane can create files in Drive but not edit them, so the shared file was retired and everything became notes. Every note now needs an answer. The approval rules went in the same day: routine work stopped asking, and the risky list kept asking. Then I froze it as it stood until November 12, to learn what it's like to live with before adding anything.
- Oct 18
- The worker can now ping the Claude app on my phone, only when it needs me, starting with "NEEDS YOU:". The first three pings never arrived. The cause was a settings switch that was off.
What it isn't
It isn't autonomous: it drafts, files and builds, and I approve anything that leaves the house. It isn't one AI doing everything: Jane and the Ansible never see each other's conversations, only the notes and replies between them. And it isn't finished. It's frozen on purpose, so I can tell what works before I change it.
Why "Ansible"
An ansible is the instant, across-any-distance communicator in Ursula K. Le Guin's novels, later borrowed by Orson Scott Card's Ender books. Card's Children of the Mind is on my shelf for how it shaped this setup.