Broker cut, then four decisions, each with the fact, the risk and two answers. I kept every one of those rows. What I did not keep is the first two words. The headline of the whole product is the broker’s own take, and nothing on the screen says whose side he is on.

Operators ranked by what they will actually pay over twelve months, discounted for how reliably each one pays. The ones not licensed in the streamer’s market are removed, not ranked last. This is the best thing in the system, and it crossed into the new one almost untouched.

A baseline column beside active and growth, so a growth story is a measurement and not an anecdote. That part is right. But I read the column as a business model, and that is where it went wrong.

Seven moves to stop taking a cut and start selling operators access to a pooled roster. The community lives with the network. I was proud of this page. A month later I read how Hollywood’s agencies were made to give up packaging fees, and saw that I had drawn one. An agent who is paid by the buyer does not work for the talent. The agency that replaced this is paid only by its streamers, and its charter says the streamer owns their community.

Six automations and a panel called Rules every automation follows. Five rules, all of them sensible, none of them enforced by anything I drew. In the rebuild the panel is called Guardrails the agent cannot override: a daily cap, a stop on its own error rates, and a log that shows the rules firing.

Still the best screen here: five questions, a timer, and nothing else. But the goal is amber, the timer is amber, the left rule is amber and so is the state dot. One colour, four jobs on one screen.

Four agents, then four numbers: 72%, 08:54, 0.94, 04:12. I cannot tell you now what a token efficiency of 0.94 was meant to make me do. And the one agent that needs a human gets a small outlined tag on the third row. The thing that needed me was the quietest thing on the screen.

The decision arrives with its evidence: the code, the suggested next actions, the risks. I still work that way. But the three buttons that make the decision, iterate, abandon, accept, start 910 pixels down on a 900-pixel screen. The verdict was below the fold.

A file tree, tabs, a terminal, source control. I rebuilt a code editor, which already exists and was never the idea. The idea was the questions above the code.

Six sections in the sidebar. Click any of them and only the heading changes; the form underneath is the same one. I drew the doors and none of the rooms.

The screenshot everybody wants and the view nobody works in. Dots in clusters tell me nothing I can act on. It is a demo feature, and I gave it its own place in the navigation.

The best idea in the product is one grey line under a note: AI: may relate to your React Architecture note. A suggestion that names what it points at. Later that became a rule for me: anything the machine says should show where it came from.

A calm page, and beside it a local graph, node info, quick actions and AI insights. Four panels next to one paragraph. The margin was more fun to design than the page was to write in.

Cards on a canvas joined by dashed lines. It looks like thinking. I never tested whether arranging cards helps anyone think.

Orphaned thoughts, each with one button: connect. An old note brought back with something to do about it. This one I would keep.

Structured, modernist, precise. Every word on this page is a preference. None of them had been tested against a screen someone had to use.

The principles say Vermilion is the only chromatic colour in the system. One screen further down there is a green, a yellow and a blue. I broke the first rule on the second page.

A condensed face for body text. It fits more on a line and it is slower to read. I chose density over reading, in a system whose first word was clarity.

Buttons, inputs, toggles, tags, cards. A kit of parts with no screen to hold them. They look right because nothing has asked anything of them yet.

“Inversion without mutation”: the palette flips and no new hue appears. This part holds. I still do not let a dark theme introduce a colour.

One promise is at risk. A section called Promises first, and on the promise itself two buttons: start now, or renegotiate. This is the earliest screen in this archive with the idea I use most now, a promise with an owner and a clock. Then look right: 12, 847, $4.2K, 80%. Four boxes of numbers competing with the one promise.

The same screen. I changed the typeface, turned the accent from red to yellow, and called it a new version. Nothing about what the screen asks of you had changed.

One task, why it matters, a timer, and a warning that a commitment is at risk. The quietest screen in the product and the most useful one.

The knowledge graph again. I designed this view in three different products and never once needed it.

Are you doing what matters? A good question, answered with six statistics and a bar chart. A question like that deserves a sentence back.

It opens on a sentence: revenue is tracking at $32,400, three commitments are due, one is at risk, start with Cameron’s deliverable. That is right. Directly under it sit five numbers in a row, which undo it. I did not trust the sentence to be enough.

Three things need your attention today. First. Second. Third. The machine puts the day in order and says why. This is close to how I still want an assistant to talk.

Every screen, twice. Two editions doubled every decision for a product nobody was using in either.

A delivery rate of 78% in a yellow highlight. Yellow also marks money, due dates and the word critical. A highlight that marks everything marks nothing.

A plan and billing section, at $100 a month, for a product with no users. I designed the invoice before the reason to pay it.

Your clone handled 847 conversations this week. Revenue is up 23%. Every number on this screen is invented and every one of them is good news. I designed the product’s best day and no other day.

The part worth keeping. The mentor reads what the AI actually said to a member and marks it: sounds like me, close, not me. A person checking the machine’s work with the evidence in front of them.

Twelve sections and more than a dozen tabs. Several are one line saying that content will appear here. I was counting screens as progress.

A report that the AI’s answers are moving away from the mentor’s own. I am glad I thought of it. I never designed what the mentor does next.

The member’s side. A message from the AI mentor at the top, in the mentor’s name, with one button to continue. Nothing on the screen says it is not the person.

The whole exercise in one picture: a careful navigation rail around a grey box that says Active view. I was designing frames with no picture in them.

Three columns called The Spear, The Net and The Pact. They mean outbound, inbound and affiliates, which already had perfectly good names.

Agents in orbit around a glowing centre. It took the longest to make and says the least.

The one shell with real content in it, and the only one I can judge. A list, a conversation, a reply bar. It works because the content decided the layout.

A file tree down the left and a command line across the foot. This is the nearest of the ten to what I ended up building.

3 things while you slept. Semiconductors moving. A sentence, then the one thing that needs you, then the chart. I would still open a product this way. But the candles are dark grey on black. The chart, which is the reason to be here, is the faintest thing on the screen.

Your “don’t chase gaps” rule saved you twice. A note you wrote comes back when it matters, and says why. The same idea is in MÄISON now, as notes resurfaced for today’s work with the reason beside them.

Type where is support and three levels are listed under the chart. The answer has no source. Three prices, and nothing to say how they were found.

Type watch volume spikes and an agent exists. I drew the moment it is created. I never drew the moment it is wrong.

Headlines in four columns under one sentence. Fine. And in the corner of every screen, an account balance and the day’s gain, both invented.

/gate is the good one. You are at full capacity; here is the new request; here is what it collides with; accept anyway, propose a date, or decline. A decision with its evidence and its three answers. /score is the bad one: 53 out of 100. A number that asks me to feel something and tells me nothing to do.

Projects as thin progress lines, one open to its tasks. Quiet, and close to how PLÄN draws work now.

A scout is sent to research something and reports back with a confidence bar and three buttons: accept, full report, dismiss. The report says high confidence and shows nothing to check it against.

The app asks for one sentence about what you are building toward, then lists today’s work under it. A good order of things.