A few people asked, so here is the real version. Not a diagram from a conference talk. The thing that runs my week.
The short answer: one PC in my office that never sleeps, Claude doing most of the work, a to-do list it reads every day, and a handful of rules that keep it from doing anything I would regret.
Some of it works beautifully. Some of it I’m still working out. Both are in here.
The PC
It is an ordinary Windows machine with sleep turned off. Claude Code runs on it all day. It can read and write files, run scripts, and use the tools I have connected: my email, my calendar, my to-do list, my newsletter and my social scheduler.
It is like hiring someone who lives in the closet and never asks for coffee.
The chats
I keep separate chats for separate parts of my work. One for marketing, sales and events. One for client work. One for admin, like bills and paperwork. One that looks after the system itself.
Each chat starts by reading a short brief that says what it is for and where things stand. When I ask a marketing question in the admin chat, it tells me so in its first line and passes the note along. I never have to carry anything from one chat to another. That one habit saved me more confusion than any tool did.
The to-do list
ClickUp is the record. Every piece of open work is a card. When a chat finishes something, it leaves a short note on the card with proof: a link, a message id, a screenshot of the live page.
“Queued” does not count. “Done, and here is how I checked” does. Something is not fixed until it has worked for real, more than once.
Two engines
Claude does the daily work and the talking. For heavy jobs, like building or changing code or a long research pass on the open web, it hands a written brief to Codex, waits for the result, reads it, and checks it before telling me anything. I get the outcome, not the play by play.
Long jobs also get limits. Each one has a done condition someone could check, and a cap on how much it can spend. If a job stops making progress, it stops, and it writes down where it got to.
The rules
This is the part people ask about most.
Nothing goes out without my yes. Every email, post or message shows up as a card with the recipient, the subject and the full text. I say yes or no. One yes covers that card, so I am not asked twice. Payments work the same way, and I type card numbers and passwords myself. Nothing gets permanently deleted without my say.
These are not polite requests. They are checks that run before each action and block it if the yes is missing. The assistant cannot talk its way around them, which is exactly why I trust it with more.
Everything else, it does on its own. Drafting, researching, checking a page after an edit, tidying the to-do list, scheduling the next post in a series I already approved. I stopped wanting to be asked about the small stuff, and I told it so, more than once.
It writes like me. Anything signed by me goes through a voice checker that flags dashes, corporate words and the tidy rhythm that gives AI writing away. If the checker fails the draft, the draft does not go out. I gave that checker away, along with a few other small tools, on the free tools page.
There is also a do-not-contact list it checks before any send. Some people should not hear from me twice.
Recipes
When a job comes up more than once, like posting an event to the listing sites or sizing an image for a specific place, it becomes a written recipe the assistant follows the same way every time. Same steps, same checks, same proof at the end. Recipes are boring, and boring is the goal. Before it builds anything new, it has to name the tool that already exists.
What works
The daily grind. Drafting replies, keeping event listings in sync, pulling numbers, writing the first version of almost anything.
The biggest win is not speed. It is that small things stop falling through the cracks, because the to-do list never forgets and the assistant reads it every day.
What I’m still working out
It asks too many questions some days and too few on others. Sign-in pages still trip it up. Some sites do not want a robot looking at them, which is fair. Long jobs can run out of room before they finish, which is why the limits and the notes exist.
And it is easy to build more system than the work needs. I catch myself doing that. A rule that sounded smart on a Tuesday can turn into something I have to tend on a Friday.
If you want to try this
Start with one rule, not one tool. Decide what the AI is never allowed to do without you, write it down, and then let it do everything else. The yes in the right place is what makes the rest feel safe.
That is also where I start with a team in my half-day workshop, using their own tools and their own work.
If you build something like this, I’d love to hear what you would do differently.