Co-founder of Japan's first philosophy consultancy.
CEO of a tech company.
Currently writing from Tokyo.
Sakurao Yoshitatsu（吉辰桜男）· Cross Philosophies (2017) · ENVLOP (2020)
Field notes from ENVLOP, a small studio that runs on AI leverage — its own products and a few client builds. The essays are where I try to work out what it means to live and work this way.
In my twenties the scarce asset was time, and I got it back by moving to the owner's side of work and keeping my costs flat as income rose. In my thirties the scarce asset is interest. It does not arrive on its own, and changing what you spend on does not grow it; it grows where you keep showing up, and it stops growing the moment you ask it to pay for itself.
Three times this year I wrote up a web3 idea from the outside and threw it away, each time because a large company had arrived first. The fourth time I went inside: a weekend hackathon and a working submission. What changed was not what I own but how fast I can read the field, which outside judgment had never given me.
Making got free. Deciding did not. A draft that costs nothing to rewrite still costs a decision every time, and a listing an AI can maintain still needs a person to decide, each week, to press send. What is left to sell is not the tool but the decisions, packed into a week small enough that someone keeps making them.
Automation gives back time rather than output, and the standard answer to a free hour is another task, so the freed capacity goes straight back into work. The more sophisticated version is building systems to manage your own idle time, which produces a rule against working by working. Automation also thins the sense that the work was yours, so the appetite for tasks rises just as the need for them falls, and no better schedule reaches that.
A sentence that needs bold type is a sentence in the wrong place: emphasis does the work that ordering and cutting should have done. So I banned bold and emoji from every document my company produces. Because an AI writes the documents, the ban enforced itself from the moment I said it — in an AI-run company, a style preference stated once becomes permanent governance. The judgment is the only part that still needs a human.
Blanks in a plan come in two kinds: facts you have not measured yet, and decisions you have not made yet. Only the first kind has a right to stay open. When an AI makes revision nearly free, an open decision costs more than a wrong one — a provisional decision produces reactions, a blank produces meetings. So the rule I now work under: every 'under consideration' gets classified as one kind or the other, and the second kind gets decided on the spot.
In July, 130 commits touched my studio's rulebook — the strategy documents, the gates, the operating principles — against 96 for the one live bet the rulebook exists to serve. Governance is the only project that never waits on anyone and never rejects you, which makes it the most comfortable place to hide. A diagnosis of meta-work as displacement, and a one-line instrument for catching it in your own logs.
When making something costs nothing, the cost does not disappear. It moves to whoever has to decide what the thing is worth. Open-source maintainers hit this first and are closing their doors; the same bill now arrives with every proposal, application and report that lands on a desk. What lowers it is not better work but the evidence attached to it — that the thing has run somewhere real, that someone checked how it fails, that a named person will answer for it.
I ran a small studio's business portfolio on the Boston Consulting Group's growth-share matrix and never measured either axis; the labels rotted into vibes, and this week I deleted them. The borrowed theory I kept differs in one testable way: I can measure its variables in my own units, and breaking it makes a sound in my logs. That is the test for an imported framework — and it matters more now that an AI will install one for you in a sentence.
I maintain three diverging copies of the same code across my sites, on purpose, against twenty years of engineering doctrine. Duplication is debt you can see; a shared abstraction is debt you cannot — coupling, billed at change time, in a currency you discover later. And the arithmetic under the whole trade has quietly moved, because the entity that maintains my copies is an AI that does not mind repeating itself. Not a rule; a repricing.
I stopped typing commands at my own studio: the tooling now checks the record and volunteers what is due, and a gate reads everything I publish and blocks what shouldn't leave. This looks like convenience. It is a transfer of ownership — of noticing itself — and the division of labor it settled into is easy to state and harder to hold: the machine keeps the state, and I keep the intent. The one hand-written file left in the operation only ever gets shorter.
Every automation is doing one of two things: transport — moving things unchanged — or prehension, the taking-in that decides what matters, whose signature is permanent discard. The obvious rule, automate the safe half, fails in your own inbox: a drafting assistant has decided what the mail was about before you approve anything. So the line is not a property of tasks. It is a door — the point where output turns into your name, past which nothing crosses unread. One door, honestly kept, buys back both the hours and the authorship; most people have either ten anxious doors or none.
Two theatre books passed through my book club: Keith Johnstone's Impro, a working manual for improvisers, and Tsuneari Fukuda's mid-century essay on why human beings need drama. They answer each other at different scales — scenes and lives die of the same thing, the attempt to author the role instead of accepting the scene — and Hamlet, patron saint of wanting to become somebody, turns out to be literature's definitive blocked improviser.
This month my book club — five of us — read Keith Johnstone's Impro, and I took no notes; within days the book had quietly rearranged how the writing on this site gets made. That kind of reading cannot be delegated: its output is not notes but a changed reader. The other kind — extraction, structure out, meaning compressed — has just become instant and free, including the abstract you are reading now. A note on what the free kind leaves out.
An operator's field note: the deliberation I use for running things ran, uninvited, on a desk purchase — through leg grades, timber species, finish types, and a two-centimeter difference between two nearly identical models. The machinery of judgment has no sense of scale; it runs at one intensity whatever the stakes. What stopped it was not discipline but arithmetic.
An operator's honest report: I run a studio by turning everything into a system, and I tried to run my own steadiness the same way. Every system I built to produce it — deleted apps, router blocks, stacked timers — made it worse; what worked was tearing them down. Subtraction, applied to the inner life.
The system that finally got me publishing was subtractive: instrument the real work, publish its trace automatically, and keep the self out of the loop. If publishing feels like a stage, the move is not to perform better — it is to build the thing so no performance is needed.
The system I built to keep my performing self out of the studio's writing worked — and then kept going: it started removing the judging self too. A system that summarizes your work is deciding what the work meant; the smoother it runs, the less you notice the deciding leaving the room. An honest report from inside that problem, without a solution.
Every document about your operation starts drifting from reality the moment it is written, because the work moves and the description does not. The fix is not discipline but architecture: keep almost no representations, regenerate them from the version history on demand, and hand-write only intent — in the present tense, with the past deleted. Running the studio with an AI agent made this non-optional: an agent believes what it reads.
I built the standard AI infrastructure — routing, retries, evaluation, orchestration — and deleted all of it within a year, because the model outgrew the scaffolding. In a domain improving this fast, infrastructure is a depreciating bet against the thing it sits on: build as little as the capability allows, and delete the moment it allows less.
A year of running a studio produced more deletion than construction — close to a fifth of every change was a removal, including whole products. When AI makes adding nearly free, the scarce skill moves to subtraction: judging what not to add, and what, having added, to take back out.
Japan's largest corporations pay for philosophy — and what they are buying is not ideas but a capacity their own institutions are built without: sanctioned room to hold a question open before converting it into a project. Nine years inside the category that grew into that gap, and a hypothesis about why the scarcity is sharpest in Japan.