How the studio runs

Most studios measure success by how many people show up. I try to measure something harder and more honest: would a real person love this enough to recommend it, unprompted, to someone they know?

That single question sits at the center of every decision. It rewards products that genuinely earn a place in someone's life, and quietly retires the ones that don't. People coming back, and people telling friends — that's the proof. Everything else is just noise on the way there.

Each product has its own dedicated agent that knows it deeply. Above them sits the orchestrator, which behaves like a shop foreman: it hands out the day's work, checks that it actually got done, and watches for problems across the whole floor.

The agents talk to each other through a shared channel — a kind of group chat for the software — so a decision made in one place is visible everywhere. I'm not in the loop for routine work. I'm there for taste, judgment, and the handful of calls that genuinely need a human: brand, strategy, and anything that can't be undone.

The cadence is deliberately simple and always identical: look back at what happened, reconcile the product's current state so today's choices rest on real facts, pick the single most valuable thing, do it, and leave the records clean for tomorrow.

Because every product follows the exact same routine, one manager can run fourteen of them at once without any single one falling behind. And a problem surfaces the same day it appears — not a week later when it's expensive to fix.

A fresh AI session usually starts from zero — no memory of what was tried, decided, or ruled out before. For long-running work, that's fatal. So every product keeps a living record of its state: what's in progress, what's been decided and why, what's still waiting on an answer.

When a new agent starts its day, it reads that record first. The result is an organization with a memory — months of accumulated judgment that compounds instead of evaporating. That's what lets one person sustain fourteen products without losing the thread.

This is a proven idea from manufacturing: a system is only ever as fast as its single slowest step, so working hard anywhere else is wasted motion. I apply it daily.

Instead of spreading effort thin across everything, the studio names the one constraint actually limiting progress — maybe a confusing first-time experience, maybe a missing piece of knowledge — and concentrates there. When that constraint breaks, a new one appears and the focus moves. It's a disciplined way to make sure effort always lands where it changes the outcome.

It's easy to say “make a great product.” It's much harder to make “great” measurable. So each product defines two things in plain terms: the exact moment a user feels the magic — the thing they'd mention to a friend — and a specific metric that shows whether more people are reaching that moment over time.

Coming back (retention) and recommending it (word-of-mouth) are the signals that count. This turns a vague aspiration into something I can actually steer toward, day after day.

Measuring love is one half; earning it is the other. So every product keeps a map of its most important user journeys and, each day, picks one to improve — cutting a step, removing a decision, swapping words for a picture, making the path to the good part shorter and clearer.

Before any change reaches a real person, it's pressure-tested by synthetic users: AI stand-ins that walk through the product like a real customer would and report exactly where they got confused. Only changes that genuinely smooth the path move forward — and I always have the final say on taste.

People — and AI — tend to grade their own work generously. I assume that and design around it. After an agent scores its own day, an independent auditor, deliberately given none of the original agent's context or motivation, re-scores it from scratch and looks for the holes.

The difference between the self-score and the cold score is measured over time as a built-in bias adjustment. It's a quiet, automatic check that keeps the whole system's view of itself honest, instead of slowly drifting into self-congratulation.

The most valuable output isn't any single app — it's the way of working itself, and that gets refined constantly. At the end of each cycle, agents reflect on what went wrong or could be smoother.

Recurring problems aren't patched one at a time; they're solved once, permanently, in the shared foundation every product draws on. Over time the studio's failure modes shrink and its strengths compound. The system that builds the products is, in a real sense, the most important product of all.

Not every task needs the most powerful — and most expensive — AI. I match the tool to the job. The top-tier model is used for judgment, design, and the decisions where quality really matters, and treated like a senior partner whose time is precious.

The high-volume, mechanical work is handed to a faster, cheaper model, and then the senior model reviews that output before it counts. This keeps quality high where it matters and costs sane everywhere else — the same way a well-run firm doesn't put its most senior person on routine paperwork.