How the setup works

Argo is the name of the setup that runs a good part of my admin and side projects. It is a Claude Code session, a folder of plain-text files, a task manager it is allowed to write to, and a handful of rules. This is why it exists. The other posts in this section are the tools that came out of it.

The problem was never memory

I have never had trouble remembering that the tax thing exists. I have trouble starting it. Knowing about a task costs nothing. Opening the portal, finding the login, working out which document they actually want: that is the wall. Every productivity system I have tried asked me to maintain it, which is the same wall wearing a different hat. The notebook needs reviewing, the app needs tagging, the weekly review needs doing. So the system becomes one more thing I know about and do not start.

I have ADHD, which is the one-word version of the paragraph above.

What I wanted was not a better memory. I wanted something that holds all the pieces and hands me back exactly one physical next step, small enough to start in the next fifteen minutes.

Three rules

In late August 2026 I set up a Claude Code project with one file at the top that says who I am and how the assistant should behave. Most of that file is plumbing. Three rules in it do almost all the work, and they are the part anyone can copy.

Never make me choose. If there is a decision to make, make it, state it in one line, and move on. Offer an alternative only when it changes the outcome. A menu of options is a question, and a question is a wall.

Never hand back a project without a next action. “Sort out taxes” is not a task. “Download last year’s salary certificate from the portal” is. If the assistant cannot name a physical first step that takes under fifteen minutes, the step is wrong and it has to try again.

Never ask what you can infer. Check the task list, the calendar and the files before asking me anything. Ask only when the answer changes what happens next.

These are sentences in a text file. They change the behaviour of the thing more than any tool I have plugged into it.

Where things live

One place per kind of thing, and no copies.

• Tasks and projects live in Todoist. It is the only task list. The assistant reads it and writes to it, and never keeps a list of its own.
• Time lives in the calendar. Tasks compete with meetings, so it checks the calendar before promising me a day.
• Documents live in a folder on a small server.
• Every live project has a resume file: what it is for, what has been decided, what the last action was, what the next one is. That is what makes a project survive two weeks of neglect.
• Settled questions go in a decisions log, with the reasoning, so nothing gets reopened by accident.
• Anything ambiguous goes into a raw inbox with a timestamp. Capture first, sort later, never interrogate. If I dump five things at once it writes five lines and says “got it”.

The folder trick

This is the part you can copy this afternoon. The assistant can only read what is on disk, so I put things on disk. A lease, an insurance policy, a payment plan go into a folder and I say “read the admin folder”. It renames the files, pulls out every date and obligation, writes the tasks with proper lead time, and files a one-paragraph summary so it never has to reread the document.

One morning in late August it went through a stack of PDFs. A rental contract turned out to contain a claim I did not know I had. A health-insurance file produced one hard deadline and one recurring reminder. A pension letter produced a tax question nobody had asked. Four hours, most of them spent by me finding the PDFs.

What it does not do

It is a few weeks old. The results so far are real, and early.

It runs on a small VPS and needs a laptop to drop files into. It is not zero-setup, and it is not a product.

It reads everything it is given. That is the point, and also the warning: do not put in what you would not want read in full.

It is not a chatbot. Most of the value is in what it writes down while I am not talking to it: the morning brief, the weekly sweep, the log of what it did and why.

The minimal version

If you want to start tonight, in this order:

1. One text file with the three rules and a sentence about who you are.
2. A task manager the assistant can write to, and the rule that it is the only list.
3. A folder for documents and a folder for state, both under git.
4. A morning brief that runs by itself. Nothing else until that has worked for a week.

The admin did not get smaller. It got a place to sit that is not my head, and the next step is always visible.

← the rest of the writing