The System
A personal operating system, run by an AI chief of staff

Not a smarter chatbot. A colleague with rules you can read.

The System is an AI that holds every obligation the founder has, runs the day, files each decision as a standing ruling and ships its own code — all under written rules that say exactly what it may do alone. Below: a real morning, replayed. Then the one rule that makes it safe. Then how to build your own.

01 — Replay

Watch a morning run

Forty minutes of real work, compressed to ninety seconds. The founder talks; the system opens the session, reads what is due, brings decisions and turns each answer into a rule it will follow next time. Click any grey line to see what it actually looked up.

ReplayA Thursday morning
    0:00 / 0:00
    • LookupThe system checking live state before it speaks. It reads; it does not guess.
    • DecisionA choice only the founder can make, briefed, with a recommended option first.
    • RulingThe founder's answer, filed as a standing rule so the same question never comes back.
    • LearningA mistake the system caught in itself, written down where the next session will read it.
    • AlertSomething that needs attention today, with the numbers that prove it.

    Names, amounts and identifiers removed; the shape of the work is unchanged. Space bar plays and pauses.


    02 — The one rule

    The one rule that makes it safe: the approval lattice

    The system may do a great deal. What it may do without asking is a much shorter list, and it is written down. Every action falls through one of three gates. Reading and drafting are automatic: nothing changes, so nothing needs a yes. Creating needs approval: it is usually undoable, but it adds weight you will carry. Deleting, spending, sending, or widening anyone's reach is always asked about, however small. A class of action only graduates to a lower gate when it passes a four-part test, and the test is strict on purpose. That is what lets you hand over real power: you never wonder what it might have done.

    • Gate 1 · GreenRead & draftAutomatic. Done, then reported.
    • Gate 2 · AmberCreateNeeds approval. Prepared, then asked once.
    • Gate 3 · RedDelete, spend, send, widen reachAlways ask. Never on its own.
    1. Reversible?Can it be undone completely?
    2. Internal?Does nothing leave the system?
    3. Objective?Is "right" checkable, not taste?
    4. Bounded?Is the worst case small and known?

    This is a 60-rule heuristic so you can feel the idea. In the system itself the same gates are a written ruling the assistant reads at every session open.


    03 — Build your own

    Four rungs. Stop on whichever one is enough.

    Nothing above needs a team or a budget. It needs a way of thinking about these AIs — memory they can check, a method they follow, rules they cannot quietly bend — and the guide teaches that one rung at a time. Every rung is a complete system on its own.

    1. Level 0 €0

      Free

      The AI account you already have, a job description and a two-minute open-and-close ritual.

      Start here →
    2. Level 1 €20/mo

      One subscription

      A paid assistant plus a free database it can open, so it looks things up instead of remembering.

      Read level 1 →
    3. Level 2 €100–200/mo

      No hardware

      Hosted agents and schedules, so work keeps moving while you are not in a conversation.

      Read level 2 →
    4. Level 3 €500+

      Your own machine

      An always-on box at home that runs builds and jobs overnight and answers only to the lattice.

      Read level 3 →

    Or skip ahead: clone the starter kit and have the obligation views and the session ritual running on a free database in fifteen minutes.


    04 — Under the hood

    Four ideas do almost all the work

    • Obligations, not tasks.Everything open has an owner and a date to check it again — so nothing rots quietly in a list.
    • The approval lattice.Three gates and a four-part test decide what happens without a yes.
    • The session ritual.Every session opens by reading what is due and closes by writing a handoff the next one can trust.
    • Rulings over approvals.Answer a question once, as a rule, and the system stops asking it.
    How it works →

    THE SYSTEM · How it works · Guide · Kit · Updates · Impressum · Datenschutz