Kaosi Unini · Producer · Writer · Researcher · About · Résumé · All cases · Print it on the press

KaosOS

a knowledge system run like a production ~ agents draft, I rule, and the record audits itself

Role
builder and only user, self-directed
When
2026-04 → ongoing
Status
in-progress
Stakes
personal
A holographic treatment of the vault's graph view, taken 11 October 2026: every note a dot and every link a line, packed into one dense cluster with tendrils at its edges and a few unlinked notes in an arc beside it. Light against a dark ground, teal at the left shading to violet at the right.
The vault's graph view in Obsidian, captured 11 October 2026, cropped and given the house holographic treatment. No note titles are shown. The author's own material.

Handed

On 28 April 2026, a scheduled task wrote the first daily note of a system built to record my working life, and stamped it with the line that names its premise: Kaosi was not present for live questions. I was working client hours between classes, but wanted to be writing fiction. The vault existed to manage everything.

It began as a half-joke: a recommender for my own thoughts, or at least for whatever I wrote down. Most of keeping your word is not forgetting it. But if my sensibilities could be encoded in a system, was using it still my thinking, or was I offloading it? It's an unnerving question, and it sits inside knowledge management. It's also the fear most people bring to working with LLMs.

A week in, an audit against a knowledge manager's handbook returned a precise diagnosis: the vault recorded well and synthesised nothing. It was a diary with good filing. The question it posed was structural ~ whether a personal archive could be made to think back.

Decided

KaosOS is an Obsidian vault, a folder of plain-text notes on my laptop, that Claude agents read and write under rules I set, with scheduled tasks that open and close each day. The system rests on a division of labour. Claude agents draft much of what lives within; nothing changes how the system works without a ruling from me. Four rulings built it.

Integrity before features. In June, files began returning shorter than they were written. The obvious suspect, a script that edited notes across the vault, made only small anchored edits; the fault lay in the sync layer between disk and sandbox, which served partial reads and accepted partial writes. Rather than patch the suspect, I bound every write to four rules: refuse anything under 80% of the file's prior length, edit by anchor, re-read after writing, and scan daily. No file has truncated since.

Audit the auditors. In September I turned the system's instruments on themselves. The morning check that reports overnight failures proved right once in four, and every false alarm traced to one cause: it inferred a task's state from a proxy instead of reading the log the task had already written. The rule became log-first ~ alarm only where log and filesystem disagree, cite the source on every alarm, and verify a repeat before escalating it. The most useful thing an agent did that month was argue against its own proposal, a loosening that would have silenced a true alarm. And after two duplicate tasks sat unnoticed for eight days, a fifth check now reads the task list itself.

Promotion by judgment. An idea once waited out a seven-day timer before it could feed a larger argument. I replaced the timer with a manual promotion gate: a flag set by hand, at a fortnightly session with myself, declaring the idea ready. Wisdom is declared by an author, not elapsed by a clock.

Four kinds of note. The first design offered a single synthesis layer. I split it four ways: an insight is noticed once, a concept recurs, a skill is a move made on purpose, and a framework is the argument that binds them. Each has its own file type and rhythm ~ nightly, weekly, monthly ~ so a passing hunch and a working method never share a drawer.

Shipped

As of 10 October: some 2,300 notes, including 165 daily notes since 28 April, 322 insights, 60 concepts, 31 skills and 64 frameworks, six of them generated by the system itself. Thirteen commands and seventeen scheduled tasks run it, and since September every new guard ships with a test fixture.

The system does work for others, too. The decision trail from which four parties build Sworn lives here, as does the script that counted the 111 decisions on the Telling the Bees page, and a 359-line seed for friends starting systems of their own.

Cut: a link scanner, retired in June.

Figure 1, One producer, many kinds of work. Top: eight rows, one per kind of work (client work, game jam, fiction, research, school, career, writing, the system itself), across 28 April to 10 October. A grey tick marks each day a stored idea was linked into that row: dense in client work through June, in the game jam through July, in school from September. Under the rows, a strip of the lessons written down each day, heaviest from May to mid-July. Bottom: three ideas on the same eight rows, a black dot for the first day each reached a kind of work. Inhabit, don't explain, a writing rule, was already in use when the records begin and reached all 8; Tone apart from content reached 7; the system's own morning-evening loop reached 7.
Figure 1. I counted from the system's own daily notes: each time a stored idea, meaning a concept, skill or framework, was linked under a kind of work that day, once per idea, kind and day. That gives 690 links, on 181 ticks (a day in a kind of work), across 80 days. On a narrow screen the same ticks are grouped by week. A link shows the idea was put next to that day's work, not what it changed there. The records begin on 28 April; an idea in use that day began earlier. The lessons strip counts the 322 insights written down. Kinds of work are grouped, and no client, collaborator or note is named. In the system's own terms, the mornings are Path A, putting what it knows to work, and the evenings Path B, writing down what the work taught.

Outcome

The measure that matters is use. The two ideas promoted to full concepts were applied to real work nineteen times between them ~ nine and a half each, against a target of two.

September tested the other half of the design. Classes began, a client delivery peaked, and the automation caught every lapse in the cadence: weekly synthesis ran one week in four, and decisions waiting for their trails rose to eighteen. The monthly report named the lesson plainly ~ the system had grown better at knowing than at acting.

A company buying AI enablement fears the same thing at a different scale, when it lets people go on the bet that what the team knew can be kept without them. My research asks what walks out the door anyway. KaosOS is that question with one participant: it remembers what I recorded and what I ruled, and the judgment is still mine to declare.

Would change

Pair every alarm with an owner from the start. Detection was cheap to automate, so it arrived first. Next time, each alarm ships with the action it implies and the person or task responsible, and the monthly report counts gaps closed beside gaps found. An alarm without an owner is a note to self with good timing.

Evidence

Reference. Every figure above traces to a file in the vault; none of these are public, and each travels on request with client and personal material removed.

Edges