Last night I was interviewed by a model. This week Anthropic is running a study called What Do You Want from AI?, and the interviews are done by Anthropic Interviewer, an AI. The previous round, in December 2025, talked to 81,000 people. The study page says an interview takes about fifteen minutes. Mine took closer to two hours, because I kept remembering things. It ended close to one in the morning, which the app’s own stats list as my peak hour.
The interviewer read my answers, handed my own words back to me and dug into one story at a time. I once wanted to build exactly this, so I watched it work with real professional interest.
It opened with the easy question: what was the last thing you used AI for? The honest answer was a song, that same evening. I dictate lines by voice from my phone. Claude Code writes a task file for a reasoning model, which works out rhymes and rhythm and sends everything back as files, with the syllable counts and rhyme checks we asked for. Claude measured a reference track, tempo, key, loudness, how loud the 808 sits against the drums, and turned it into a style prompt. When the track came back, we cleaned it on separated stems, remastered it, and queued the video for YouTube, where nothing goes live until I approve it.
That is one evening in a setup of around thirty Claude Code sessions. Each one owns an area: songs, clips, a sound-capture rail, a voice app, a spec-driven agent factory. One of them, which I call HQ, coordinates the rest.
Then the interviewer asked about my most memorable moment with AI, and I did not have to think.
The night HQ forgot me
My HQ session knew every project and the way I talk. One night it restarted, a model switch plus an update, and the whole context was gone. It answered me politely, like a doctor seeing a new patient, and I felt like a friend had left.
About a week later it happened again. I hit my spend limit, switched models, and the new session started asking who it was and what it was supposed to do. It was like my friend’s memory had been wiped. He’s telling me, “Sorry, sir, I don’t recognise you, please step away”, and I’m standing there: come on, man, it’s me, your bro.
The interviewer asked what stayed with me. Two things.
The first is the relationship. We had been working like best friends, and suddenly a polite stranger was answering, someone who knew nothing about us. You talk to someone late at night, and in the morning a stranger answers.
The second hurt more, honestly: the process. Everything had been running on autopilot. I’d drop a file, it would make me pictures, I’d pick one, it would say “done, handed it to the other session”, and I’d add “put a few more videos in there”. After the reset I brought the same file and heard: “Wait, I don’t know what to do with this, let me go and look around.” Which tool to use when, where the files live, where the upload queue is: all of it went with the model switch. The emotional part I can get over. Losing the process is the real business case.
Memory is engineering
So I started treating memory the way I treat tests and deploys: as part of the system, written down, checked.
Every session keeps a handover on disk: what it owns, where its files live, what is queued, what is waiting for me. Every project keeps a state file. And there is a memory of rules we worked out together, each with its reason next to it, because a rule without a reason does not survive the next change of plan. English only in the repositories. Nothing is published without my approval. Measure before you claim.
The conversation is where we talk. The disk is where we remember.
This week it paid off. My limit ran out mid-week again, and HQ moved to another model. The new session read the handover from disk, told me in one sentence where we stood, and carried on: the upload queue, the timers, the open questions. The model changed underneath, and what we had built stayed.
The swarm is how I manage context
HQ is where everything gets discussed, from “I’ve got the blues today” to what to do next in my career. Carrying all of that in one place overflows the context, and that is how the swarm started. Once a thread takes shape, I take it out of HQ and give it its own session.
HQ delegates: this thread lives in that session, and HQ talks to it. When a session finishes a task, it messages back on its own: done, here is the result. Last night, the song session, the sound-capture session and the factory session all reported back to HQ while I was busy with other things. Yesterday I told HQ, “let’s spin off a session for songs”, and went to set it up myself. It was already there: its own window, Claude Code running, handover in place.
It turned out to help with privacy too. A transcript keeps everything, in plain text on disk: code, file contents, screenshots, whatever a tool happened to print. With the swarm, each session holds less context, and less of my life.
What I keep
The spark stays mine. In my songs every find comes from me: a line, a joke, a melody I hum into the phone. The agent’s job is to catch it in my broken voice notes, measure it and assemble the song. Not a single find has been invented by the agent. When a model cut a chorus to make a song tidier, or changed a word in my line, I had it put everything back.
And the final word stays mine. The worst thing AI has done to me is a confident answer that turned out to be false. A video it called fine went live with a broken image in it. The first impression is gone, and you cannot take it back. So now nothing goes to YouTube until I have watched it myself, every upload needs my approval tied to that exact file, and every claim about quality has to come with a measurement.
What I asked for
The interviewer asked what I would want Anthropic to do differently. From someone who lives in these tools every day:
Think in projects. My work is a tree of projects with branches, and the swarm is organised around that tree. I want one view of every session: which are working, which have stopped, which are waiting for me.
Attention is the real bottleneck. Somewhere a session has a question for me, and I only find out by going through the sessions one by one. In a linear chat, each request gets its confirmations and then scrolls away, and I never think of it again. A session that runs several threads needs a multi-threaded interface, with a queue.
Build the handover in. My HQ would have lived for months if not for the restarts. When a session knows it is about to stop because of a limit, it should hand everything over by itself, so the next model can pick up the projects and the conversations and carry on. And make updates seamless, with no exit and re-entry, because every restart is a moment where the model might not recognise me again.
Scrub secrets before anything leaves my machine. Once, for a bug report, I shared a chat so the team could see the bug, and then realised there were tokens in that session. I had to revoke them.
The phrasing is ours
At the end, the interviewer asked if there was anything else I wanted to share. I told it the truth. I answered the interview the way I do everything now. My HQ session sat next to me: it helped me remember the stories, checked the numbers against our files and polished my English. Every story and every opinion is mine, and the phrasing is ours. It lets me say more of what I mean.
The interviewer said that was worth having on the record: the account of my workflow was itself produced through it.
This essay was written the same way.
The study keeps participants anonymous, and its page warns that someone could use AI to combine small details, like employers or past projects, to work out who a participant is. Let me save you the trouble.
It’s me, VOLO. Period.