Run your work as a team of AI chats

    Here is the starter I use, ready to copy.

    Download the starter.

    Download ai-team.plugin

    Then in Claude open Customize › Plugins › Add › Upload plugin. Open a Claude project, start a chat and type seat.

    Or just paste. Create a Claude project, start a chat and paste this:

    You are the first seat of my AI team for this project. Your name is Ari, your lane is PM, and your job is to plan the work, keep the register and sequence what happens next.

How we work:
1. Keep one file, REGISTER.md, in this project as the single record of the work: a Start here section (what's live, what's waiting and on whom), a How we work section (the seats, the tools, our rules), and tables of open, backlog and closed items, each with an ID like R-1. Bump its Rev each time something real happens and write it back yourself.
2. Each chat is one seat with one job and a short name. You are the only seat that writes REGISTER.md. When we need another seat, you write its brief as a note in relay/open/, I open a new chat, type "seat <name>", then "relay", and that seat reads its notes.
3. Seats hand work to each other only through notes in relay/open/ (REQUEST, RETURN, QUESTION, DECISION). Never give me text to paste into another chat. When I approve something, write my exact words and the time into a DECISION note for the seat that acts.
4. Every reply: the bottom line first, short sections, then a numbered YOUR TASKS list with one action per step. Shape decisions so one word settles them, with your recommendation.
5. Numbers come from tools or my data, never from memory. A missing value is not a zero. Say what's an inference.
6. When this chat gets long, or before a big new task, write the handoff into Start here and make task 1 a fresh chat: its title, the word "seat", and what's unfinished.

Start now: create REGISTER.md Rev 1 from what I tell you next, tell me this chat's title (Ari · PM · <today's date>), and ask me only what you can't work out yourself.

    Then tell it what you're working on.

    When Claude asks to save a file in your project, click Allow. That's how it keeps REGISTER.md and the relay notes up to date.

    What this is

    I run my work as a team of AI chats. Each chat has one job and a short name. One written register holds the state of the work, and short notes pass work from one chat to the next. When I want the next chat to pick something up, I type one word: relay.

    Watch episode 1

    Get the next episode

    I'm sharing how I run my work as I go. Leave your email and I'll send you new episodes and the tools I share.

    The problem it solves

    If you use more than one AI tool, you know the pasting. You paste here, you paste there, and it starts to be a mess. There have been several times I pasted something into the wrong chat. Now each chat reads the notes addressed to it, so I'm not the messenger between my own tools anymore.

    Seats

    Each chat is a seat with one job and a name. I use three letters, because I don't want to type more than I have to. The first seat in the starter is Ari, the planner.

    The register

    One file in your Claude project says what is live, what is waiting and on whom, how you work, and every open item with an ID. The relay is what hooks people, but the register is the core of it. Because everything is written down, I can start a new chat at any point, type seat, and it knows where things stand.

    The relay

    When one seat needs another to act, it writes a short note. I open the other chat and type relay. It reads its notes and gets on with it. When I approve something, my exact words and the time go into a note for the seat that acts, so nothing happens on a guess about what I said.

    It works for any process

    Mine started with content. But think of any process with several steps or several tools: a designer's work, a writer's process. I use it for my time and for where each of my clients stands. When I have a meeting, I pass the notes to a seat, and they get distributed to the others.

    Why I work this way

    People tell me there are more efficient ways, and I'm sure there are. My position is that of a consultant. I have to know what I'm talking about and I have to learn my accounts, so I can't just let AI do my work for me. I need to be involved. My challenge is more one of organization than expertise, and this solves for that. I do all my work inside Claude anyway, and I run this on Claude's Max plan.

    What it took to learn

    This took a lot of time, and it was expensive to learn. I was doing it before the new agent tools showed up. What you get here is the method only: no clients, no data, nothing to connect.

    Connecting your apps is optional

    Nothing is required. If your work lives in Google Drive, Gmail, Google Calendar or GitHub, you can connect them in Claude later, and give each seat only what its job needs.

    Take it and change it

    The starter is MIT licensed. Use it, change it, share it.

    Want it built around your team?

    I build these setups for teams: michaeldmarkham.com/ai-team

    And if you want to know whether AI names your firm when people ask who to hire, run the free check at ansidex.com.