8 Systems, 3 Weeks: Building My Personal AI Operating System

Claude opened early access to Fable, its most advanced model, for a limited window — about two weeks when it was first announced, then extended by one more. Fable's back to a paid tier now: Max plan or pay-per-use, and pay-per-use is not for the faint of wallet.

By the time I understood how tight the window actually was, I decided I'd squeeze every token out of it. I'd already made a decision I didn't fully register until later: I was going to spend all of it on infrastructure nobody could see.

.

Here's the question that took over everything else that month:

Not a product. Not content. Not anything with a link you could click.

My answer was direction — and a foundation solid enough that everything I build after this gets easier.

Fable is good at untangling complicated, ambiguous problems. The repetitive execution work then gets handed to Sonnet, which is faster and doesn't need the same depth of thinking to do it well. That division of labor is the part I'd actually recommend to anyone working across multiple models: use the expensive intelligence for direction, the cheaper intelligence for momentum.

The trade-off was real. Three weeks with no product shipped, no revenue, nothing to point to when people asked what I'd been working on. Just eight systems that, from the outside, looked like I'd lost the plot a little. (Claude spent a fair chunk of those three weeks telling me I was over-engineering. Probably fair.)

What I actually built

01 — KOS (Knowledge Operating System).

A central store for every fact and piece of context across my projects. Not just a digital junk drawer — it defines how information gets filed, connected, and reasoned about, so knowledge compounds instead of piling up. I dump raw notes into an inbox and it sorts itself out. The failure mode is obvious and I've hit it already: garbage in, garbage out.

02 — NEXT.

A task-sequencing assistant that answers one question every morning — what should I actually work on today — and interrupts before I drift into something that feels like work but isn't. Does it stop me from chasing shiny distractions? Not really. Turns out the system was never the problem.

03 — PEN.

My content system. It holds a model of how I actually write — my opinions, my go-to phrases, the words I'd never use — so drafts carry real substance instead of sounding like generic AI output. It pulls from KOS and whatever's current in NEXT, so a draft isn't starting from nothing. This post is one of its drafts — and so is the proof it works: PEN rebuilt the same story for a Claude power-user community I'd been quietly reading for a long time, and that version did unexpectedly well, far outperforming every other version of the same story. More on how that actually works in a future post.

04 — MiroFish.

Decision support for the calls that actually keep me up at night — money, business bets, life stuff I keep flip-flopping on. I write out whatever I'm stuck on, and it works the situation from three angles: the numbers, the practical execution, and the uncomfortable question of whether I'm actually solving the problem or just numbing a symptom. It ends with a recommendation, the reasoning behind it, and the specific conditions under which that answer would change — advice with an expiration date built in, instead of a one-time verdict I'd forget the logic of a month later. (Idea sparked by the YouTube channel Everyday with Captain — I heard the framework there and built my own version of it, full credit.) The gap I haven't closed yet: no system to check whether what it told me to do actually panned out. That's next.

05 — Life OS.

Connects the long-term plan to whatever I'm actually building day to day, with milestones and scheduled check-ins on whether the direction still holds. Most useful when you work alone and the people around you have politely stopped asking what you're up to.

06 — FKE (French Knowledge Explorer).

Tracks lessons, homework, and my teacher's corrections, then quizzes me daily on whatever I keep getting wrong. The real gap was never not knowing the rule — it's knowing all thirty-ish French verb conjugations and still picking the wrong one in the moment. English gives you six.

07 — Wealth Tracker.

This one started selfishly. I had no clear picture of my net worth, and I'd taken a 50%+ loss on Thai equities without fully registering how deep it went until I saw everything laid out in one place. Having that visibility is what let me recover deliberately instead of guessing. It worked well enough for me that I opened it to a small group of early users. They're still using it, feedback has been strong, and it's what this website is actually built around.

08 — WithPuii.

This site — the home base pulling everything I work on into one place. Had Fable help design the structure, since the work spans finance, language learning, marketing, and photography. The common thread across all of it, once I stopped and looked, is the same: helping people make clearer decisions with less noise.

The actual cost

All eight systems took close to 60 hours combined over three weeks. My eyes and wrists have filed a formal complaint. Work-life balance left the building the moment I went all-in on this — more desk pain, less sleep, the usual toll of building alone.

What I got back wasn't output. It was a framework I can actually see — I know what leads to what now, and why, instead of circling the same questions by myself at 1am.

Whether that's worth it depends on what you're optimizing for. For me, right now, it was.

.

If you're building something similar — whether it's a knowledge system, a decision tool, or just trying to get a clearer financial picture like Wealth Tracker started as — I'd genuinely like to hear what you're working on. That's most of why I write these down.