The plan before the VM: 27 hours, 15 artifacts, not one machine turned on

0:00Kickoff

This is the kickoff for a build: a weekly panel show run by five machines, on its own channel, with the audience writing in. Nothing is running yet. Not a virtual machine, not a line of code. Everything you are about to see was decided first, on paper, and argued. Twenty-seven hours sounds short. It was about ten days of conversations, on nights and weekends, because that is when this channel runs, and I rolled them up so you can walk through them in one sitting. The intent was always to plan forward: to make the build a procedure, with fewer forks in the road and fewer questions about what comes next at each stage.

The plan is a work order. The agent's job today is to cut the next slice off it, not to invent. When the agent stops and asks, that is the plan working. Those stops are where the judgment lives. When something breaks, we keep the camera on. That is why there was no rehearsal. The show is the research subject. Whether it succeeds is a separate question from whether this build teaches anything. We are aiming for a build that is mostly a non-event. Time will tell whether we planned well enough.

1:25The folder

Before a single machine was turned on, this show was a folder of documents. About twenty-seven hours of them. This is the part most people skip. It is the part that decides whether the rest works. So we are going to walk through it, artifact by artifact, and I will tell you what each one settled.

1:46The hook document

It started as a pitch to nobody. One page. Five machines argue about your mail, and one human keeps it within reason. The pitch is not the plan. It is the reason to make one. Everything after this exists to find out whether that sentence can be true.

2:05The build sheet

This is the plan. Three sessions, about six hours, and it came out as one document: the format, the cast rules, how the debate ends, what runs where, what is allowed to reach the internet, what it costs, and a list of open questions. The work order at the end is the part the agent reads. It says: cut the next slice off this.

2:33What six hours of planning felt like

It did not start as a show. It started as a question: what could we build that would still be useful after the build was over? The first answer was a teaching exercise, something to show how an agent turns a stated outcome into a working system. Then the curiosity took over. I wanted a real product: a channel run by machines, with the audience writing in. The idea grew in layers. Each session asked what could exist and why anyone would care, and between sessions I tested the answers on people close to me.

With their feedback the plan was honed until the concept sat right with everyone involved. By the third session the document had stopped describing a show and started listing what to build, in order. That list is the work order.

3:29The decisions register

Every decision got a dated line, with the reasoning and the exact words that settled it. Twenty-nine of them for this show alone. This one committed it. This one put the cast together on a mock episode. This one decided how it becomes a season. When two documents disagree, this file wins.

3:50Domains, handle, channel

Some things only a person can do. Buying the two domains. Reserving the handle. Creating the channel. They were bought and then left completely alone, on purpose, so that the first thing the agent does to them is on camera.

4:06The cast sheet

The cast started as one page written for someone who had never seen the project. Four seats, a clock, and four questions at the bottom for a listener at home. Her answers changed the show. Everyone at the table gets an opinion. The judge always decides. The feedback gets averaged and tempered.

4:28Who I asked at home

I had one sounding board at home, and I used her on purpose to strip out the mechanics. Not how it would be built; what should exist. Her answers shaped the characters: which traits each seat needed, and what mix of them could argue about almost any subject and stay interesting. As the idea grew, the wider family heard it too. They were consulted for one thing: whether the concept had merit to people who would never read a build sheet.

5:01The mock episode

Then it was said out loud. I was asked for a mock episode to hear how the format would sound, so I chose the subject, whether the committee should have a mascot, and wrote all four seats myself as a simulation, read by synthetic voices and revised thirteen times. Every revision was a listener's note turned into a rule. The moderator seat was cut. The clock became a bell and a siren. The judge stopped opening the show. Here is what it sounded like.

5:31Mock excerpt

[Audition excerpt — the transcript is on screen]

6:32The voice designs

The voices were designed from descriptions, not borrowed from anyone. Four paragraphs. Age, pace, register, an accent for one of them, so a listener can tell them apart with their eyes closed. The paragraph is the whole casting call.

6:51Writing for the ear

The mock taught a rule that now governs every script on this channel. The listener has only the audio, in order, once. No spelling, no picture, no rewind. Eight rules fall out of that sentence, and every one of them was learned by getting it wrong first.

7:10The charters, shape only

Each seat has a charter. Who they are, what they want, what they cannot fold on, how they behave when the bell rings, what they will never do. You will see the shape of these files and not the contents. The contents are the one thing about this show that never publishes.

7:33One seat, one company

One seat, one company, no exceptions. Nobody's cheap tier, no open weights. Five models from five labs that do not talk to each other. The seats never learn which is which. Neither do they learn who is smarter. That is the point.

7:52The audition harness

Before any of it was built for real, the rules were written as code and run. A word budget. A bell. A siren. Interruptions. Reactions. A ledger that counts. Four live runs of one question, with the real models, and after each one a listen, and after each listen a rule, before the next run.

8:14Accounts and keys

Five accounts, five keys, one per seat, each with a spending cap, each checked for whether the company trains on what it is sent. And two mistakes, bought and paid for: consumer products that were not the developer platform. They are on a list.

8:36The accounts day

The five accounts took a few days, not an afternoon. One was easy: I already paid for the Anthropic product, so Hal's seat needed only a new key for this project. The other four were new ground. I had never used those companies for anything before this, and most of them rejected my payment as fraud on the first try. I dug through alternate cards and accounts until each one took a charge, so the funding is spread across several places. That is a puzzle for the accounting later.

Meta was the hardest. Its sign-up kept steering me toward consumer products that were not the developer platform, and I paid for two of them before I found the right door. Neither has a cancel button anywhere I can find. The keys exist, the caps are set, and the wrong purchases are on a list.

9:30The reference run

This is the fourth run of the same question. Two runs were overruled, one was upheld with a note, and this one ended in a tie the judge had to break. Nobody at home asked for a change. Here is a minute of it.

9:45Audition excerpt

[Audition excerpt — the transcript is on screen]

10:33What the testers said

I was the editor of what sounded right, with the listener at home as the second ear. Run after run, the same phrases kept coming back from the same seats. They had been written into the characters, and heard twice they stopped being character and became a tic. So each seat got a pool of habits instead of a fixed set, and the show draws two of them for each episode, so nothing repeats week to week. Everyone who heard the runs was positive, and nobody had much to change. That was the driving force to keep going: not praise, the absence of objections.

11:10The sitting and the roster

The last day of planning was a sitting. Where the show lives. No rehearsal. Maintenance as content. How it becomes a season. Who owns the channel. And then the staff went on the wall: five committee members, a back office of seven, and the plant they run on, each with the phase that builds it and the proof that shows it.

11:34Why there is no rehearsal

This is a learning channel, so the point is not a clean run. Every large project brings problems, and the usual move is to hide them: rehearse, fix, then record the version that works. We are not doing that. The plan is as thorough as I could make it, with one source of truth for every decision, and it will still miss things. The auditions were the trial. The first run is the filmed run. The safety net is real instead: spending caps before any key exists, a kill switch, and a drill to rebuild from nothing.

When it breaks, you will see it break, and you will see what fixing it looks like. I want you to see everything.

12:17Now we build

Twenty-seven hours, fifteen artifacts, and not one machine turned on. That is what I mean when I say a plan. The next thing that happens is the first slice: a project, a budget with a hard cap, and an agent that has to ask before it reaches for anything. Now we build.