Home / Personal operating system / How to build one
Build

How to build a personal operating system

Most guides to this start with structure, which is the wrong end. Structure is the part people enjoy designing and the part that matters least. Build it in the order the layers actually depend on each other, and you can stop at any point and still have something that works.

A personal operating system has four layers, and they are not equally important. Each one is only worth building if the one below it is already true. The usual failure is building the second layer beautifully and never building the first, which produces an elegant, empty structure that you fill in by hand until you stop.

Layer one: it files itself

Before anything else, the system has to receive the world without you retyping it. Email, calendar, meeting transcripts, and the things you read all have to land somewhere on their own. This is unglamorous and it is the layer everything else rests on, because a system fed by hand is a system that stops the first week you are busy.

The concrete version: a folder of plain files, an assistant with access to your mail and calendar, and the rule that anything arriving gets filed to the project it belongs to rather than to an inbox you triage later. The mechanics are in automatic meeting notes and AI email triage. Build this and nothing else, and you already have more than most people's systems deliver.

Layer two: structure, kept deliberately thin

Now the structure, and the discipline here is to build less than you want to. You need projects, people, decisions, and reference. That is close to all of it. The elaborate taxonomies people build at this stage are the ones that go stale, because every extra distinction is another judgment someone has to keep making.

NodeWhat lives thereThe one rule
ProjectsStatus, open questions, decisions, next actionOne file is the truth; every view is derived from it
PeopleHistory, what you owe them, what they owe youThe relationship lives here, not in the deal
DecisionsWhat was chosen and, critically, whyThe why is the whole point, and it is what nobody records
ReferenceWhat you read and want to keepFiled by what it is for, not by what it is about

If you want a starting layout to copy rather than design, there is a template, and comparisons of the standard structures live in the PKM methods roundup.

Layer three: the routines run by name

Once material arrives on its own and has somewhere to go, the recurring work becomes worth automating. A weekly review, a morning brief, a follow-up sweep, a status pass across every project. Each of these is a routine you describe once and then invoke by name rather than reconstruct every time.

This is the layer where the system starts saving you real hours instead of just holding your files, and it is also where most people would happily stop.

Layer four: the agentic layer

The last layer is the one people mean when they search for an agentic personal operating system: the system doing the reversible work before you ask. Drafting the reply you will obviously need to send. Preparing the brief for tomorrow's meeting overnight. Noticing that a decision you parked three weeks ago is now live again. It presents the work and waits for your approval rather than acting unilaterally, and the distinction matters more than anything else in this layer.

Build these in order and every stopping point is a working system. Build them out of order and you have a structure with nothing in it, or an agent acting on a record that is not true.

Can you build this with ChatGPT?

Partly, and the limit is worth understanding before you start rather than after. A chat assistant is genuinely good at layer one's judgment calls and at layer three's routines. What most chat interfaces cannot do is read and write the files on your own disk directly, which is what makes layers one and four real rather than a conversation you then copy out by hand. If the assistant cannot open your folder, edit a file, and save it, you are still the integration layer, and you will get tired.

That is the reason this site's build uses Claude Code rather than a chat window: not because the model is better at thinking, but because it has direct access to a folder of files you own. The comparison in detail is at Claude Projects versus Claude Code, and the broader tool question at Notion AI versus Claude Code.

What to skip

Skip the perfect naming convention, skip the tag taxonomy, and skip any layer of structure whose only justification is that it feels tidy. Every one of those is a maintenance cost you will pay weekly for a benefit you will not notice. The thing that makes a personal operating system last is not how well it was designed. It is that nobody has to keep it alive.

The full build, in order, hands-on

The book walks all four layers with the prompts printed, one working piece at a time. No code, and you own every file at the end.

Get the book
Same four layers, company scale

See it already built

A working system running the same layers on a company's email and meetings, with the agentic layer switched on. Try the demo.

See the demo