For a year I ran my life and my businesses from a system I had designed myself. I called it HQ. It worked. And it was exhausting.

Eight AI agents with names. Dozens of surfaces. A dashboard for everything. Every morning began with the administration of the thing that was supposed to save me from administration.

So I started again, from one sentence.

The person operates intentions, not the database.

That sentence is the whole product. The AI reads, files and connects. The screen shows the one thing that matters now. Everything else sits one gesture underneath. Old software with a chatbot in the sidebar is the thing MÄISON refuses to be.

I designed it between Thursday 1 October and Monday 5 October 2026: thirty-four working screens, a constitution for every word on them, a permission model for three AI colleagues, and a build plan in eleven phases. I wrote down every decision as I made it, including the ones I reversed the same afternoon. The whole ledger is here.

There is no password. There never will be. You give an email or a phone number and the house sends a one-time link, good for ten minutes on this device. Or you touch a passkey. The first version sent a six-digit key. It lasted two days.

Look at the foot of that small window. --read --login --presse. The options are written as flags, because that is what they are. And that small window is not a login form. It is the terminal from the first room of the house, at a smaller size: the same three squares and the time, the same toolbar, the same dot face for what you type.

I did not design that on purpose. I noticed it.

What happens in those nine seconds

It is a geometry problem, not an effect. The red moment holds for 700 milliseconds, for everyone the house knows. The window’s lines fade in 160. A copy of the window is placed at its measured position and animated, left, top, width and height, over 420 milliseconds to exactly where the terminal of the next room sits: centred, 1180 by 248, computed with that room’s own layout arithmetic so the handoff is exact to the pixel. The field settles from #000000 to #0e0e0e. The toolbar fades in at the foot. The room assembles around a terminal that is already in place.

About 1.3 seconds, door to table, and there is no frame in which two different objects are on screen.

Clients and contractors get a plain fade to their own page. Their pages have no terminal to grow into, and motion with no referent is decoration.

Four doors in one Saturday

The door you see is the fifth. On 3 October it changed four times, and the log has all of them.

• An editorial red poster, with a masthead, captions and four lines on who may enter.
• Red, the name and a terminal, nothing else. The masthead, the captions and the footer went.
• Black, with red as the founder’s ritual. The log says it exactly: red is not spent before anything has happened. The door went black; the screen turned red only when it recognised me.
• Red for everyone. The last change that day. The log again: red is the moment of being known, for anyone the house knows. A client gets the same 0.7 seconds I do.

Two days later it learned to greet you and ask, and the exit I had built and liked, the window sliding down to become the command bar, was replaced by the window growing in place. The old one made the bar the hero. The new one says what is true.

The first thought is typed large, in the dot face, eighty-four pixels and shrinking as the line fills. Press Enter and the window widens, settles at the foot of the screen, and the conversation opens above it. The same object, again, changing size.

There are four ways to be answered, and you choose: explore (one question deeper), decide (options and their cost), vent (just listen), learn (what I expected against what happened). More than one colleague can sit at the table. @märcel gives one of them the floor while the others wait.

The AI may speak first. It may never interrupt.

This was the decision that removed the most. A colleague that needs me, to clarify, confirm or choose, opens a thread with one question and waits. It shows as a count beside a word and one line under the window. It is never pushed. There is no status bar and there are no notifications anywhere in the house. What needs me lives in places I go to on purpose.

Provenance is not a feature

Anything the AI says about my world has its sources one step beneath it. The mark is two words: why ›. Under the sentence on the home screen. Under the decision at the top of the day. Under every answer a client gets. Extreme compression at the surface; unlimited inspection underneath. I can only delegate to something I can check.

Nothing scrolls, and here is the arithmetic

Scrolling hides the end of things. You never know how much is left, and a long page quietly becomes a feed. So every screen in the house is a whole. Declaring that was easy. Keeping it once the data grew was the design.

The grid reads its own top edge. Its height is the window, minus that edge, minus 84 for the command bar, minus 40 for the pager. Columns are ⌊(W + gap) ÷ (minW + gap)⌋, rows likewise, with a minimum size per kind of card: projects 300 × 220, canvases 300 × 240, seeds 380 × 280. With fewer cards than room, a second pass chooses the arrangement that keeps a card’s shape. For each possible column count it scores |ln(card ratio ÷ ideal ratio)| + 0.5 × empty cells, and the lowest wins. The weight on empty cells is heavy on purpose. The half-filled bottom row was the thing I could not stand.

What does not fit is in pages, or one level deeper. A laptop gets fewer rows. A tall monitor gets taller cards. Nobody gets a scrollbar.

What it cost. For a day I let shift-and-scroll turn the pages. It worked, and it broke a gesture that already meant around. Withdrawn the same day. Two gestures, two meanings.

That session is the part I would defend in any room. Each question arrives alone: ? 1 of 3 · approve · märcel, the question, a note, and how to answer it. Yes or no. Done or later. Per car, or a monthly fee. My answers stay in the log with the colleague’s acknowledgement. And when the last one is answered, the project says: Thank you. That was everything. We are done for today.

Then it asks nothing more that day. Software that knows when to stop asking.

One rule here looks small and is not. A meeting exists only inside a project. I can schedule one from the calendar in a sentence, meeting with Priya fri 3pm about the ark offer, but if the sentence names no project the calendar asks which one. A meeting that belongs to nothing cannot be made. Every meeting therefore has one link, which is its invitation, its room and its record.

No bell. No badge list. No inbox of notifications. The only thing on this screen competing for my attention is the truth about my day, set in type large enough to read standing up.

Two morning notes from the colleagues lived on this screen at first. They moved to the conversation room, where they can be answered. This screen has one job.

This surface was first called Ättentes, awaiting. On the fifth day I renamed it Dëcisions, on the desk and on the phone, because everything in it asks for a yes, a no, or a word. The name should say what you are there to do.

The rule that shapes it comes from Kenya Hara, and it is the first principle in the design document: one thing in full ink. Every screen has one thing that matters most, at full white or in red. Everything else steps back into grey. If two things are in full ink, one of them is wrong.

Piëces: writing, with colleagues who write too

Piëces is where I write. On the fifth day it became where the three colleagues write as well, and that changed what it is.

The colleagues keep the procedures they work from here: how to judge a dealer’s website, who qualifies as a lead, what to say on the call. Each one carries a version log, and every version carries its evidence. Version four of the website procedure added a finance-calculator check, because twelve of the fourteen dealers who booked had none. Version three dropped page speed, because in sixty-one scores it never once changed a decision.

That is the design rule for everything the AI produces: it shows its work, and then it shows the work behind the work. My own pieces are never revised without asking.

Lëcture: emptiness that scales

A reading shelf is lovely with ten books and unreadable with a thousand. The first shelf had one row, and it was built for ten.

A new canvas starts blank. gather, followed by a goal, reads across my books, my notes, my writing and my projects and places what bears on it around the goal. Each window says where it came from. A red line joins it to the goal.

A post on the canvas is a master. One gesture makes a render per medium, essay, newsletter, carousel, short. Edit the master and the renders are marked drift, with a way to regenerate. Every made thing carries a manifest of what it came from.

And one principle I wrote down early and kept: honest stand-ins. Where the prototype cannot reach a real service, it says so on the screen in one quiet line. The images on this canvas are dot-matrix placeholders and are labelled as such. A prototype that fakes a result as real has taught you nothing.

Where it came from. Two objects. The red TrackPoint in the middle of a black ThinkPad keyboard: one point of colour on a machine otherwise entirely about work, and you always know what it is for. And Braun’s ET66 calculator, where Dieter Rams allows one yellow key among the greys, and it is the equals sign. The colour marks the act, not the mood. MÄISON’s accent is literally ThinkPad red.

The options. Red as a theme: the first door was a red page, and it was handsome. Red as emphasis, sprinkled on anything important. Or red as one state with a strict set and a strict clear. I built the first two and lived in each. The theme spent the signal before anything happened. Emphasis reproduced the problem of every dashboard, where the eye learns to ignore the colour because it is everywhere.

So I wrote the state machine first, in one sentence: a thing is red while it waits on you, and only the act that answers it turns it black.

The mechanics. An email, text, call or voicemail carries needs while it is unanswered, unopened, missed or unheard. Opening does not clear it; a system that turns black because you looked is lying to you by Thursday. Then the area, by three failures. Whole page red: body text on saturated red fatigues inside a paragraph. Body black but the reply box red: the reply is the cure. Final: red on the 56-pixel bar and the title band, transition: background .4s ease, and nothing else. White on this red is about 4.5 to 1, right for labels and deliberately wrong for paragraphs, which is one more reason the body stays black.

What it cost. Three reversals in one afternoon, and an hour of red sender icons with white dots that were correct in theory and noisy in practice.

A person, on one page

A contact card stores how to reach someone. A relationship lives in everything else: the thing they mentioned on a call, the promise you made in a meeting. So the person’s page is built as ammunition for the relationship, and every line of it is sourced.

Six panes and no scrolling. Every exchange with him is one gesture in, where the navigation bar itself becomes the kinds of exchange: all, email, texts, calls, meetings, documents, notes. One more level is the exchange in full. Out is the same gesture reversed.

The gift ideas are grounded in what he actually said. A colleague can arrange one. But only after I say yes, because spending always asks.

Search that shows how it understood you

Every search in the house is an object resolver. It finds the thing, or it offers to make it. And where the question is in plain words, it shows its reading of the question before it shows the answer.

A year ago I designed a system with forty-six agents. Every feature had its own. It was impressive, and it was dead. HQ had eight. On the third day of this week I folded all of them into three. A fourth was proposed and scrapped the same day.

The tell that I had been wrong before was simple. With eight agents I had to remember who did what. With three, I just talk.

The roles are the permission model

This is the part I am proudest of. You can tell what each colleague is allowed to do from what it is for. Säinte knows everything she is allowed to read and changes nothing: she proposes. Läurent acts, but only inside a boundary that was approved. Märcel decides within a ceiling and brings me whatever is above it. I did not bolt a permissions system onto a cast of characters. The cast is the permissions system.

Five laws, in code and not in prompts

• Money past a ceiling, signatures, publishing in my name, deleting anything: ask me.
• Consent. Calls only where consent allows. An opt-out on any channel stops every channel.
• Honesty. Each speaks as itself or as a business with an approved card. Never as me. Never as any real person. It says it is an assistant when asked.
• My calendar. Meetings are booked into hours I have opened, never over protected ones.
• Privacy. Health data never leaves my own view.

A prompt is a request. A hook before every tool call is a law. The design specifies the second.

Abilities are taught, not hired

The first version let me hire an agent from a sentence. I changed it to teach a capability: a described ability becomes a skill of one of the three, a file I can read, or a mission if it has a schedule. Plain machine work, finding leads, verifying them, transcribing calls, runs as unnamed assistants: lowercase names, no inbox, no voice. Processes, not people. The people list always shows three.

The detail that shows the thinking

Läurent can sell by phone. An AI voice on a phone call is, in law, an automated call: the FCC ruled so in the United States in February 2024, and Canada’s rules require express consent for a synthesized voice. So the design says his voice goes, by default, only to people who said yes. Cold lists are opened by email or text where the law allows. And the consent ledger is built before the dialer. That is a design summary and not legal advice, and the document says so at the top.

And when a campaign has no list, he does not wait. Finding is reading, not reaching out, so he builds one himself and reports before anyone is contacted. The specification gives the report he should write: 640 found · 412 verified · 188 may be emailed · 0 may be called yet. That last number is the design working.

What I needed was not an agent framework. It was three persistent people: each with a character, a memory, skills it has been taught, sessions it remembers and a schedule it keeps. Hermes Agent gives each of those as a profile out of the box: a SOUL.md, memory, skills, sessions, cron. Three profiles in its official container, in the cloud, beside the record, always reachable. That is the reason in the build plan, in eight words: persistent people out of the box, always reachable.

A colleague’s character is a text file I can open and read. I wrote the first draft of each. Märcel’s begins: You speak little; when you speak, it is a decision, a reason, what it costs, and what you found.

The models are Claude, by API key, one per colleague, set in a hidden settings room. The runtime and the models are separate choices on purpose, and the plan keeps a managed version of the runtime as an option if hosting it myself becomes a chore.

Decision 0

Before any code, one sitting, one table. Every row has a reason, and the reasons are short.

Voice is the one thing that never leaves home. Calls, voicemail and the glasses are transcribed on a machine in my own room. The settings screen says it in four words: never leaves the house.

A phone call needs an answer in under a second, and a full agent loop is too slow for that. So a live call runs in three parts. Before it, Läurent writes a brief: who, why now, what we know, the one goal. During it, a fast voice layer works from only that brief, with three tools: check the calendar, book, hand to a person. After it, Läurent takes the recording, updates the record, scores the call and writes the note. The slow mind prepares and reviews. The fast one talks.

It begins with a fiction, because a voice needs somewhere to come from. Imagine a design bureau on the rue de Sèvres and its sister office in Montréal, both founded before the war, both working in two languages without comment, which in 1981 quietly built an operating system for their own use. Minitel is on the desk. The typewriters are Olivettis. Nobody has ever written an exclamation mark in a memo.

From that, three registers, and one rule for each.

The umlaut

Every word of the house’s own vocabulary carries exactly one. If the French word had an accent, the umlaut takes its place: ÉTUDE becomes ËTUDE. If it had none, the umlaut sits on the first vowel: MAISON becomes MÄISON. One per word, never two.

It is not French and it is not German. It is a mark. And the reason is in the type: the umlaut’s two dots rhyme with the dot-matrix face. The name of the product is spelled in its own typeface.

The reversal

By the third day, French was on everything: every tab, every filter, every status. It was charming and it was slow. On the fifth day I wrote the rule that fixed it, and rewrote a glossary of fifty labels in one pass. Töus became everything. Ättente became awaiting.

The test is one sentence long, and it ended a hundred small arguments: if a word names a place you can go, a tool you can open, or one of the three, it is French. If it tells you something, sorts something or asks you something, it is English.

The French makes the house a house. The English means you never have to decode an instruction.

The faces

Three typefaces, each with one job. Array, a dot-matrix face, for the one thing on a screen that should be read from a distance. Söhne Mono for everything read or typed. Courier Prime for the private journal and nowhere else, because a journal should feel like a typewriter. There was a serif. It retired on the second day.

And the languages underneath

TypeScript, end to end, for the web app and most of the phone. Kotlin, in small modules, only where the phone’s hardware demands it. SQL for the record. And Markdown for the two things that matter most: a colleague’s character is a SOUL.md, and everything it has been taught is a SKILL.md. The most important files in the system are written in plain sentences, and I can read all of them.

Each change was written as a small, ordered transformation of the screens, so the design can be replayed like a log. The prototype is built by a generator, and it went from generation 123 to generation 387, staying interactive through all of them: working links between rooms, a working door, live replies, the keyboard grammar, the motion. Every change was checked in a real browser at several sizes before it stood.

The design ends in documents an engineer, or an AI coding agent, can build from. A build plan in eleven phases. A specification for the three colleagues. The language constitution. A build spec for the phone. And a one-page rulebook that ends with a list titled Never.

What is proven, and what is not

The design is complete and the prototype works. Open it and use it; nothing on this page is a mock-up. What it cannot show yet is the thing that matters most for an operating system: a month of real life run through it.

So the rebuild is planned in the order that tests it soonest. First the record. Then the day. Then the projects. Then the colleagues, who arrive only when there is a house for them to live in. Every phase ends in a gate, a sentence that is true or false: you open MÄISON in the morning and it tells you, truthfully, whether anything is yours today. Not true, not finished.

Three questions are open, and I am watching them: whether Säinte’s findings stay true enough to trust, whether the colleagues’ freedom holds inside the laws, and whether red stays rare as the product grows.

How to read it

Each line is a choice: a rule adopted, a thing removed, a name changed, an earlier choice undone. The reason beside it is the one my design log gives, shortened. Each line ends with the document and line it came from.

A line that says gut instinct, a hunch or by eye has no written reason. I am not going to invent one now. Taste is part of the work, and so is saying when that is all it was.

Struck lines were reversed. The numbered ones, D-153 to D-194a, are product decisions; the rest come from the dated design log. Early entries use the names of the old cast of agents, which folded into three on the third day.

The starting point

Buying a phone on Marketplace and selling it for more is a real business, and it has a ceiling. The price is whatever one person can haggle, every deal depends on that one person, and nothing compounds: the hundredth phone is as much work as the first.

I did not start with screens. I started by asking what the business actually sells, and the answer I proposed was not a phone.

EBN sells an ownership cycle, not a phone.

Buy the old one. Sell the replacement. Buy that one back in two or three years. A flip happens once; a cycle comes back. Everything below follows from that sentence: the price engine, the pre-orders, the cash-out screen, the buyback promise on the receipt.

A Sell says: I’ll sell this phone at this price, if it is what I told you. A Buy says: I want this spec at this price, and here is my money on hold. A Bid is a reseller’s standing offer for twenty units. An Upgrade is a Sell and a Buy joined by one condition: swap my 13 Pro for a 15 Pro if I pay $300 or less. A deal is two orders matched.

The second idea is firmness. Every order sits on one of four levels: watching, confirmed, accepted, completed. Only some of them count as money, and only an accepted order can set a price. It sounds like bookkeeping. It is the thing that keeps the business honest about demand: one buyer watching five models counts once.

And EBN is a dealer, not a marketplace. It buys every device and sells every device; customers never trade with each other. That one decision means every phone passes through one inspection and one ledger.

The offer is the best place the phone can be sold, minus everything it costs to get it there. Five lines: the best exit, a margin by price band, a risk buffer by grade, an allowance for testing and fees, and the cost of the handoff. For an iPhone 16 Pro in good condition with a live reseller bid of $845, that is 845 − 90 − 25 − 20, which is $710 if the customer drops it off.

Each exit is priced on its own terms. Local retail pays more and takes six days. Wholesale pays less and takes three. A reseller’s bid pays today. So every exit also carries a holding cost per day of waiting, and a slower exit has to pay more to win.

Three guards

• No number without fresh data. A price is firm only with enough recent sales behind it. Otherwise the page says we’ll confirm by text within the hour instead of guessing.
• A floor. No deal is offered below a minimum contribution after the handoff is paid for.
• A swing check. An offer that moves more than 8% from yesterday waits for the owner’s approval instead of going live.

And a rule I would put in any system that reads outside data: nothing fails silently. A broken price source turns red and keeps the last good prices. It never zeroes them.

The cash-out screen unlocks once the appointment is verified by both sides, before the handoff. The quoted cash is shown as the down payment on whatever comes next. Four paths: keep the cash, take a certified used phone, take a new phone on a carrier plan, or type any model and have it sourced. The customer arrives with the next phone already chosen, and the upgrade happens in one visit.

Every path is shown as one all-in monthly number, with tax and fees included. Carrier pages rarely add it up for you. This one does.

Cash before inventory

A buyer can pay for a phone before EBN owns it. The payment is authorised, not captured: the money is held, and it is only taken once the phone has been sourced. If sourcing fails inside seven days, the hold is released and the buyer was never charged. It takes inventory risk and working capital close to zero, and a paid order is the strongest signal of demand the engine has.

The wait is honest. The page says delivery tomorrow only when a matching phone is already in stock or in the pool of quoted sellers. Otherwise it gives the real estimate.

A field agent meets the seller and takes eight photographs, every one on a white mat: front, back, edges, a white screen, two settings screens, the signed-out account, the seller’s ID. An AI reads them and returns a grade and a confidence. Then the deal is green or amber.

Green pays itself. The grade matches the quote, confidence is 90% or higher, the price is under the auto-pay limit, the payee is confirmed. Amber waits for the owner, who sees the photos with the AI’s marks and three buttons: approve the new price and pay, keep the quoted price with a reason, or call.

The rule underneath is the one I would defend first: no payout without an inspection record. An agent cannot change a price, a grade or a payee. The customer sees any new price and the photos before anything is sent, and can walk away.

The AI has a job description

One text number, answered around the clock. It can quote a phone, ask for photos, offer times, and book the handoff. It never promises a final price before inspection, never goes above the engine’s offer, never confirms a payment, and never argues. Its price is the engine’s price. When a seller wants more, or someone is upset, or it is not sure, it hands over with a one-paragraph summary and goes quiet on that thread.

Texts only go to people who agreed to them, and stop works instantly. The consent ledger is part of the data model, not a setting.

The name. EBN, Electronics Buyers Network. It says what the company is and what makes it different from one person with a Marketplace account: a network of buyers behind every offer.

The colour. Ink, paper, and one signal: a sharp lime, used as a fill and never as text. It marks the number you are being offered and the thing that just matched. A second colour, a burnt orange, exists only for risk. If you see it, something needs a human.

The type. Geist for everything people read. Geist Mono for every number, every time: prices, grades, IMEIs, dates. A price set in a monospaced face reads as a figure from a ledger, not a figure from a pitch.

The photograph. Every phone on white. It is an inspection rule first, because the AI rejects any photo that is not on the white mat, and a brand rule second. The storefront and the evidence are the same picture.

When EBN sells a phone, its inspection, its price history and its owner are already in the system. So three things become possible that a flipper cannot offer. A certificate: inspection record, warranty, proof the data was erased. A guaranteed buyback: a written minimum price for twelve months, printed on the receipt. And a public price index, built from real sales, that becomes the place a city looks up what a used phone is worth.

Put together, the offer is one sentence: trade in, pay the difference over time, guaranteed buyback when you’re ready for the next one. A phone subscription without the subscription.

None of that is in the first build. It is phased, and it starts only once the engine has fifty completed deals behind it. I wrote it down now because it changes what the first build has to record.

What “working” means

I did not want done to be a feeling. The spec opens with nine sentences, and each one is either true or false on the day it launches.

1. Every offer can be explained.
2. No number without fresh data.
3. No payout without an inspection record.
4. Every open deal has exactly one next action, and one person who owns it.
5. Demand is counted in people, at the right firmness.
6. The money balances. Contribution is computed, not typed.
7. Texts only go to people who agreed to them.
8. Nothing fails silently.
9. A seller can get an offer on their phone in under 60 seconds.

Behind those sits a test plan of more than thirty scenarios, written before any code: two customers trying the same slot at once; a lead arriving at 11pm; a phone that drops a grade at inspection; a payout that bounces. Launch happens when every box is ticked on production, with real phones and real texts.

What I cut

What is proven, and what is not

This is a design and a specification. Nothing here has been built yet, and nothing has met a customer. Every number on every screen is sample data, and the screens say so.

The pricing rules are starting points. They have to be calibrated against the business’s real deals, and two questions still block the first hour of the build: which twenty models it trades most and what it actually paid and got for each. The parts that touch law, such as texting consent and record-keeping, are written as design rules and need a lawyer’s answer before launch. The spec says so in plain words.

What I can show is the thinking: a buy-and-flip taken apart to what it sells, rebuilt around a record, a price you can explain and a promise on a receipt, and drawn for every person who would touch it.

The brief

ARK Business is a holding company with three operating companies — Marketing creates demand, Sales converts it, Operations makes the business work better — each with its own partner and its own buyer. One shared website would have had to say everything to everyone. Four unrelated websites would have looked like four companies.

One family, four arguments.

The design problem was a family resemblance strong enough that a dealer principal moving from one site to the next knows he is still talking to ARK, with each site free to make the single case its buyer needs to believe.

What each site has to prove

I started from the visitor, not the layout. For each company I wrote down the one sentence a buyer should think after thirty seconds, then designed the site as the evidence for it:

Site	The visitor should think	So the site…
ark.business	These people understand companies.	Unfolds like a document: three companies, three industries, one film on how ARK thinks. The words carry it.
arkmarketing.ltd	These people have taste — and know why the work exists.	Inverts the order: work floods first, words interrupt. One paragraph, three nav items, the archive as a sideways contact sheet.
arksales.ltd	These people know how to convert demand.	Opens on a timer — 00:00:08 — because speed to lead is the product. Shows the follow-up text they'd send against the one most dealers send.
arkoperations.ltd	They can rebuild the machinery under the business.	Has to feel engineered itself: a changelog, a status line, a real search; case studies told as operating systems.

A method for borrowing well

For every site I took three reference studios the partner admired, wrote down what each did best — not what it looked like — and took only that. For Marketing: Basis's work-before-words, Alright's native-ratio image river and flat type hierarchy, ET Studio's one-paragraph homepage and contact-sheet archive with its 0 / 348 counter, because volume is the argument. Each design document opens with a table of what was taken and from whom, so the borrowing stays honest and the reasoning survives the designer.

The fixed foundation

Four voices, each with a territory. Excon for display, never below 40px; Switzer for interface, never above 32px; Erode for anything read; Sometype Mono for numbers, labels and buttons. Between 32 and 40px is a dead zone: nothing lives there, because two sans-serifs at 36px read as a mistake rather than a pairing. The hierarchy is done by the jump from Excon 700 at 72px to Erode at 17px — wide enough that a page still reads with every colour removed.

One edition, nine tokens. No night mode. Near-white paper with a green cast you feel rather than see, white surfaces that lift off the page, and a sage accent split into three values because one mid-tone stops reading as colour at 10px. Emphasis is a sage highlight behind ink, not coloured text. One colour breaks the family — clay — and only for errors, because a green-tinted failure message reads as success for the half-second that matters.

Rules you can lint. Zero radius, no gradients, no tenth token, tabular-nums on Switzer wherever it touches a number. Each rule ships with a checklist, so a contractor can verify a page without me in the room.

The industries as an architecture

The content structure carries a business argument. ARK's three industries — Automotive, Real Estate, Commercial — are a ladder of larger asset markets: things people drive, things people own, things that produce money. Each one is financed (car and loan, property and mortgage, business and acquisition financing), so the same machinery compounds. The sites show Automotive as active and the other two as future focus, which makes the roadmap legible on the homepage instead of hidden in a deck. I argued for Commercial over Finance (too narrow) and Business (“ARK Business — Business”).

The fifth system: software

Products are not websites. For ARK's software I wrote a separate system — ARK OS — that keeps the family's discipline but changes what a tool needs: light and dark, three type roles (editorial numbers, IBM Plex for detail, Plex Mono for data), an icon rail, status colours with real meaning, and a page structure that reads editorial at the top and functional below. It became the system for the dealer and broker operating systems in this portfolio.

What I cut

First version	What it became, and why
A previous system with a yellow accent and a mandatory day and night edition	One edition, one hue. Two editions doubled every decision for a website nobody reads at midnight.
A visibly tinted paper	A tint at a tenth of the strength. Visible green read as “eco brand”; a faint cast reads as paper.
A single sage value	Three — fill, line, text — because sage either fails as text or disappears as a hairline.
A shared homepage for all three companies	Four sites. Each buyer needs a different first sentence.
A tinted dark theme for the software	Achromatic neutrals with sage on the primary action only. The tint made a tool feel like a lifestyle app.

The problem

Dealership software is built around the car: inventory, deals, paperwork. But a dealership doesn't make money from a car. NOTT Autocorp, a Winnipeg dealer, earns six ways — owned retail, consignment, finance and leasing, service, detailing and wraps, an EV experience — and the real money is the customer who comes back for five of them over years. No tool they used could see that customer. Each department saw its own slice.

Design the business before the screens.

So I started with a model of how NOTT earns, then designed one record that every role reads from — the rep, the sales manager, finance, the service writer, the general manager, the owner — each seeing their day, not a permission-trimmed copy of someone else's.

Principles

Every role gets a day, not a dashboard

The rep opens on Tess's day: appointments, the new lead waiting, deals in play, and a note from her manager — “reply to every new lead in under five minutes with two appointment times.” The sales manager sees the floor and the desk. Finance sees files waiting on documents. Service sees repair orders and dollars open. The owner and GM see one thing: what needs our attention today, sorted critical, important and below the fold, each with who owns it and a button to ping them.

Promises are measured, not hoped for

The system is built on commitments a customer can feel: a person replies within fifteen minutes; The Three — a best fit, a best value and a wild card — arrive within forty-eight hours of asking for help. When a promise slips, it surfaces on the owner's screen as “Marcus T.'s Three is overdue by 3 days”, not in a monthly report.

The customer, across every department

The customers view follows a person through the journey — service, lead, appointment, purchase, first service, detailing, the next change — with lifetime gross and every department they've touched. It turns the flywheel from a slide into something a manager can sort by.

Training that matches the business

New reps learn the business before they learn a script. The Academy opens on how NOTT makes money and the clock on the lot — a 2017 Audi R8 that has sat 140 days has cost almost $11,000 to hold. Then role-play: AI customers built from NOTT's real inventory — an EV-curious commuter, a Porsche owner who wants to consign, a first-time buyer who needs financing, an out-of-province buyer for the R8 — each with a goal, and a score on discovery, product knowledge, honesty, the next step and the flywheel. A deal desk lets a rep see what a deal is actually worth once finance, protection and the trade are in.

The pitch: change the model, not the city

Ziggy Motors' owner told us two things: getting people through the door was hard, and he was thinking about moving to Vancouver. I didn't pitch software. I wrote a page that took his problem apart from first principles — the showroom is a cost, not the product; low traffic is a symptom; a move changes rent and rivals, not the reason — and proposed a different model: four lanes instead of one front door. Express (“I found the car”), Concierge (“help me choose”), Explore (“just looking”, members not leads) and Seller (their car becomes inventory without Ziggy's cash in it).

The move became a reversible sixty-day test — sell into Vancouver from where he is, read cost per funded deal by lane, then decide. And a calculator showed the change in a year's gross after holding costs when cars sell faster and more of the lot is consigned. The ZIGGY OS demo, adapted from NOTT, sat at the end of the page as proof the model could be run.

What I cut

First version	What it became, and why
A dashboard of charts for the owner	A list of what needs attention today, with owners and actions. Owners don't want to interpret; they want to act.
Separate tools for sales and service	One record, seven views. The money is in the customer who crosses departments.
Training as product knowledge	Training as the business model first. An advisor who knows how the store earns sells differently.
Pitching software to a dealer thinking of moving	Pitching a model, with the software as proof. The question he had wasn't about tools.

The prototypes run on sample data shaped from NOTT's real business; the next step is connecting them to the dealership's actual records.

The starting point

My partner did not start from nothing. When he took me on he already had a working business: employees, payroll, contractors and a delivery system. He grows streamers’ communities and places them with sponsors, and he is good at it: one of his streamers went from four people in chat to four thousand. But the business had hit a plateau, and he brought me in to help it scale.

A month ago I designed the software for that business, a broker’s operating system. It is in the archive now. A better tool for the same business was not going to move it. A different kind of firm might.

He wanted the first serious talent agency for streamers, the way CAA is a serious agency for actors, and he wanted it to look as if it had been there for fifty years. He runs the agency day to day. My part is a minority stake and the systems I have built for it.

Two constraints shaped everything. The agency had to be clean by design, in a sponsor economy known for hidden deals and opaque contracts. And it had to be remote: no office, no local shoots. So the software is the office.

I had one day. I spent the first part of it on the business, not the logo.

The analogy that almost misled us

The first instinct was real estate. Streamers as agents, viewers as buyers, sponsors as houses, the agency as the broker. It was a useful frame until I pulled it apart. In this market the sponsor is the buyer, not the house. The scarce thing is the streamer’s audience and its trust. That makes the business a talent agency, not a brokerage.

The difference is the revenue model. A broker can take a margin from both sides. An agency cannot, without becoming untrustworthy. Hollywood already ran this experiment: the writers’ guild forced every major agency to give up packaging fees from studios by 2022, because being paid by the buyer is a conflict with the talent. We started where Hollywood was forced to end up.

It also killed my own idea. The old system ended on a page called From broker to network: pool the roster and sell operators access to it. Read against this rule, that page is a packaging fee with a nicer name. It did not survive the day.

Three inheritances

I studied three institutions and borrowed from each on purpose.

• CAA, 1975. Five agents left William Morris and designed a culture before they had clients. No names on doors, no private client lists, pay tied to the agency’s results, and two rules: be a team player and return calls promptly. I took the team structure and the shared information. I left the management by fear.
• Ogilvy. Research before creativity. He trained at Gallup and wrote down what his agency learned. So research is a department here, and every clip and sponsor read is tagged with its result.
• Sterling Cooper. The pitch as theatre. The signing pitch became a produced moment: a sixty-second reel cut from the streamer’s own footage, delivered over video.

The charter

The strategy condensed into nine commitments, in plain words. They are public, and every later decision was checked against them.

1. We are paid only by our streamers.
2. Full deal visibility. The streamer sees every term of every offer.
3. Real money only. Sponsorship is disclosed on stream.
4. Licensed operators only.
5. Adults only.
6. Responsible-gambling messaging as standard.
7. No predatory terms.
8. The streamer owns their community and their likeness.
9. Exit freely.

A charter is easy to write and easy to ignore. The rest of this page is what it looks like when the software, the website and the outreach are built to keep it.

Three reasons. Surnames would tie the clean agency to my partner’s other business by name. Consulting is the wrong promise to make to talent. And CAA kept names off its doors deliberately, so that clients belonged to the firm and not to a person.

The answer came from the two of us: Kennedy & Hargrave, after two streets that matter to us. It has the cadence of a partner firm without a founder on the door. It also carried two coincidences that went into the brand story. Kennedy is the era Mad Men opens in. And a Thomas J. Hargrave ran Eastman Kodak, the company Don Draper pitches the Carousel to.

I searched for existing businesses with the name and recommended trademark checks in Canada and the United States before anything is printed.

The direction is editorial and Swiss. My partner dislikes serifs and loves Söhne. He asked for colour from “the German Olympics”, which meant Otl Aicher’s identity for Munich 1972: the first fully systematic Olympic design programme, led by a bright sky blue and white. Aicher left black and red out of his palette as the colours of power. He also co-founded the Ulm School of Design, the school widely seen as the successor to the Bauhaus. That lineage became the backbone of the brand.

Three rules, straight from Aicher

• Black is for type, not for fields. The dark theme is deep Munich blue, not black, for the same reason.
• No red anywhere. Not even for the live dot. Orange does that work.
• Fields are flat. No gradients, no tints of tints, never two accents fighting on one surface.

Every pairing of text and field was checked for contrast: black type on the bright fields, white type only on dark green, violet blue and deep. The hex values are close screen approximations of the 1972 standard, and the kit says so. They are not official references.

Display type is semibold, tight and always flush left. No serifs and no italics: emphasis comes from weight, size, colour or position. Corners are square everywhere except the live dot and portraits. The whole identity lives as tokens for two themes, nine text styles, a spacing scale and a voice guide with words to use and words to avoid.

The grammar is short enough to memorise.

• A 24-unit square grid.
• Solid fills and no outlines, with a two-unit gap where a line would normally be.
• Edges only at 0, 45 and 90 degrees.
• Circles only for heads and dots.
• Bars 2.5 units wide.

So the inbox is an envelope whose flap is separated from the body by a diagonal gap. The lead agent is a target with a 45-degree signal and a square marker for found. Deals are two interlocking diamonds. Operators are a building with square windows.

These are original drawings in the spirit of the 1972 system. Aicher’s own pictograms are licensed artwork, and I did not copy them.

I rebuilt the shell as a Swiss poster. The wordmark sits top left. Three pictogram buttons sit top right: call, text and notifications. A narrow numbered rail runs down the left, and the active room is filled in its colour. A strip in the same colour runs across the top of the page. Then the page’s one headline figure, set at poster scale in black on a full-width field.

No cards. Content sits in open sections under a two-pixel black rule with small mono labels. Tabs are underlined, not boxed. Status labels are flat blocks of colour. The colour of a room follows the division that owns it: blue for the workspace, orange for new business, green for talent and deals, light orange for reporting.

I did not rebuild the record. The deal matcher, the baseline every streamer is measured from, the decisions that arrive with their evidence: those were right in the old system and they came across whole. What changed is whose side the record is on.

The lead agent, and the rules it cannot override

The agency runs an AI agent in the background. It reads public channels, finds streamers who fit, writes to them, and puts every reply in front of a person. I designed its control room: a funnel from found to booked with the conversion at each step, a queue with a fit score and the reason for it, the three-email sequence, a live log and a pause switch.

On the sample data the funnel reads 1,240 found, 212 qualified, 188 emailed, 23 replied, 9 positive, 4 calls booked. The poster number is not the 1,240. It is the 9: positive replies this week. That is the number a person has to act on.

The part I care about is a panel titled Guardrails the agent cannot override.

• At most 60 first emails a day, from the agency’s own domain, spread across working hours.
• Only to business addresses a streamer publishes for enquiries, and only about representation. No casino offers, codes or links in outreach.
• Every email names the agency and a person, gives a mailing address and carries a working unsubscribe. A stop on any channel suppresses that person everywhere.
• Adults only. A channel whose creator or audience reads as under 18 is skipped and never stored.
• New or edited templates wait until a person says yes.
• It stops itself if bounces pass 3% or unsubscribes pass 1%.
• Every reply is handed to a person. The agent never books, promises or negotiates on its own.

The guardrails are design, not decoration. They are written to keep outreach inside Canada’s anti-spam law and United States email law, and they are what lets the agency say clean by design and mean it. The log shows them working: Suppressed 2 addresses after unsubscribe requests. Skipped 3 channels whose audience reads as under 18.

One inbox

Agent replies, email, every text and every missed call land in one inbox, from known contacts or not. The agent labels each reply Interested, Question or Not now. Texts are threads, with the agency’s side on Olympia blue. Missed calls and voicemails are tagged, transcribed and called back from the same screen. A number that replies STOP is locked everywhere.

That is CAA’s second rule, return calls promptly, turned into a room. A missed call is a decision somebody needed.

Four roles, one record

The same system is four different products depending on who signs in. The owner sees commission, risk and the decisions that need him. An employee sees their own work. A streamer sees a portal about themselves. A prospect sees one thing: a way to apply.

The top row is the brand kit: cover, logo suite, colour, typography, pictograms and applications. The second row is an exact capture of every room in the Office, plus the dialer and the text composer. The design system is installed on the board, so the colours and text styles are available to anyone who edits it. It is for my partner today and for whoever joins us next.

The work is scattered across a white field around one headline: Talent and creative agency for streamers. On your side of the table. You drag it, use a trackpad, or press the arrow keys. The page never scrolls. Tab moves from piece to piece and the field glides to each one.

The departments are Munich colour blocks. The charter is a typographic poster. The growth story is a light-orange field that says 4 to 4,000. The Office appears as real screens. And between them are the people the agency is for.

• Detail sheets that always fit. Picture on the left, short text on the right, Next and Back for anything longer. If the window is too small for the copy, the type shrinks. There is never a scrollbar.
• Filters. All, Agency, Studio, Office, Brand, Streamers. The site opens on Agency, with everything else faded back.
• Chrome in black and white. The four navigation buttons carry the colour: green for About, orange for the Charter, violet for the Office and solid black for Apply. Blue stays reserved for the ampersand.

The people on the field

An agency site needs faces. So five clips and six stills sit among the departments: podcasters, music streamers, DJs, gamers, essayists and cooks, each in the kind of room that work is made in. The clips play in place and do not open. They are atmosphere, and the pieces around them are the argument.

We worked through kh.tv, kh.studio, kh.house and others before my partner found khtalent.haus. It works on three levels. It is short. Talent says what the agency does before anyone clicks. And .haus points back to the Bauhaus, the school whose successor Aicher co-founded. The address and the identity share the same roots.

What I cut

What is proven, and what is not

This was made in one working day. I want to be exact about what that produced.

What exists: a business model and a charter, a name, a brand kit, an original pictogram set, a working prototype of the Office, a working prototype of the website, and a board that holds all of it. You can open all three and click through them.

What I am not showing: the business has clients, employees and contractors. The roster and the team are under NDA. So every name and figure in the Office here is sample data, and no client is named on this page. These prototypes are not connected to a real inbox, phone line or platform account. The application form checks what you type and does not send. The outreach rules are design rules, and they need a lawyer’s read before the agent sends anything.

Where the design falls short of its own rules: the no-scroll rule holds on the website and not yet in the Office, where the long rooms still scroll. Söhne is not licensed yet, so what you see is its stand-in. The Munich colours are screen approximations. The trademark checks are recommended, not done.

Eight hours is enough to decide what a company should become and to make that decision visible. Whether it takes the business past its plateau is not something a design can prove. That part comes next.

What I would point to

• A business model expressed as design. Paid by streamers only is visible in the charter, in a bar chart in the Operators room, in the agent’s guardrails and in the first paragraph of the site. Trust is the product. The design makes it checkable.
• A historical system used with discipline. Munich 1972 supplied a palette and a grammar. I kept the rules that came with them and drew everything new.
• Software that looks like the brand. Poster-scale numbers, numbered pictograms and a colour for each division make an internal tool feel like the firm’s own printed matter.
• My own work, thrown away. The best idea in the old system was the wrong idea for this one, and it went into the archive the same day.

The loop

I don't start with screens. I start with how the money moves — the six ways a dealership earns, the cut a broker takes from a cheque, the day a founder actually has — and I write it down until the model is clear enough to argue with. Then I find the promises inside it: the things a customer or a colleague could feel if they were kept, and measure if they weren't. A person within fifteen minutes. Three cars within forty-eight hours. Nothing published without a yes.

Only then do I design the record — what the system has to remember so the promises can be kept — and only then the screens, which are the record seen from one role at one moment. I build the screens as a working prototype so every decision can be felt, not just seen; I live in it; and then I cut. Most of the work is the cutting.

Business → promises → record → screens → live in it → cut.

Instincts I trust

These are not rules I read. They are judgments that kept turning out right, each with the moment it did:

One colour means one thing.	MÄISON has one accent, red, and it means this waits on you. When I let red spread — a whole email page, every history row — it became a mood. I pulled it back to the bar and the title. Information is the smallest area that still can't be missed.
If you have to interpret it, it isn't designed yet.	The dealership owner's first screen was charts. I replaced it with a list of what needs attention today, each line with an owner and a button. He doesn't want to read a trend; he wants to ping Nadia.
Rank by what is true, not what is advertised.	Casinos were ranked by headline commission. A high rate from a house that pays late is a low rate. The deal matcher ranks by twelve-month payout adjusted for reliability, and unlicensed houses don't appear at all.
A name is a place; everything else is an instruction.	The house's rooms are French with one umlaut; everything that tells you something is plain 1981 editorial English. The test is one sentence long, and it ended a hundred small arguments.
Scrolling hides the end of things.	Every screen in MÄISON fits the window and pages when there's more. It forces a decision about what matters now — the thing scrolling lets you avoid.
Borrow the behaviour, never the look.	For each ARK site I took three studios the partner admired and wrote down what each does best — a one-paragraph homepage, work before words, a contact-sheet archive — and took only that, in a table, so the borrowing stays honest.
Emptiness is capacity.	A shelf that is lovely with ten books is noise with a thousand. The answer wasn't a smaller shelf; it was rows that show what fits and count the rest, and a library one level in. Kenya Hara, applied.
Kill the demo features.	The broker's reference had a custom cursor. It looked wonderful for thirty seconds and slowed real work. Gone. A permanent red practice screen: gone. A fourth AI colleague for documents: gone. A gesture that paged and navigated: tried for a day, gone.
Show the work's work.	The colleagues in MÄISON carry sources, versions with evidence, and why it resurfaced. A broker's decision arrives with the statement that caused it. Provenance isn't a feature; it's what makes delegation possible.
One object, door to table.	The login window in MÄISON is the terminal you will work in, small. When the house recognises you it grows into place. The first second should feel like the rest.

Reversals, with evidence

Taste is mostly knowing when you were wrong quickly. Three of the day's reversals, before and after:

A ledger, not a mood board

Every product decision in MÄISON is written down with its date, and its reason where I had one — 193 numbered so far — so the design can be replayed and argued with. The whole ledger is public. A dozen, verbatim in spirit:

D-172	One door, no passwords. The field turns red for 0.7 seconds when the house recognises you — everyone, not just the founder.
D-172a	The door asks first: read or login? An address or a phone gets a one-time link. The six-digit key is retired.
D-172d	The door's window grows into Ëtude's terminal; founder and partners land at the table. Clients and contractors fade to their own page.
D-178a	No leads, no waiting: Läurent builds his own lists before he asks for one.
D-179a	Säinte researches every day and leaves one finding, with evidence and an experiment, before the founder wakes.
D-181 withdrawn	Posting to two personal sites from inside the house. Integration I didn't need; the sites are kept by hand.
D-184	Names in French, everything else in 1981 editorial English — with a glossary of fifty labels rewritten in one pass.
D-185a	A person's sheet fits one page, without a stage track; the record of every exchange is one level in.
D-187	No frames drawn over the screen. The dot type is the cue.
D-189b	In Cöurrier, red is the bar and the title. The body and the reply stay black and white.
D-190a	SOPs the colleagues work from carry a version log with the evidence for every change; the founder's pieces are never revised without asking.
D-193	Lëcture holds a library: three rows to the width of the screen, the rest counted; a lettered, paged index one level in. Thousands, without noise.

The ledger is how I keep taste honest. A decision I can't write down in a sentence with a reason is a decision I haven't made yet.

What this means if we work together

You get someone who will ask how the business makes money before opening a design tool; who will write the promises down and build the record that keeps them; who will prototype in code so the team can feel the decision by Thursday; who will cut his own work without being asked; and who writes his reasons down so that anyone on the team can argue with them.