The restructure

For most of 2026 the product’s headline was five modes. A capture could become a note, a quick answer, a long think, a built document, or a piece of research: five surfaces, five explanations, an onboarding flow that taught you all of them. It was the thing I was proudest of, and it was the problem.

The diagnosis

Count what a new user had to hold in their head before the app made sense: five pages, five modes, recognition, PARA buckets, wiki topics, briefings. Roughly fifteen novel concepts, against a working-memory budget of about four.

The root cause wasn’t the number, though. It was what the concepts were. The information architecture had given system processes their own rooms (here is the inbox, here is the wiki, here is where briefings live) when the user never arrives with a process in mind. They arrive with one of three questions:

• Where did that thing go?
• What needs me right now?
• Where does everything stand?

An inbox is not a question. Neither is a wiki. Both are answers to my problems, promoted to furniture in someone else’s house.

The trade

So the surface came down. Five pages became two tabs: Today for the time axis, Life for the structure axis, and every other view became a lens on one of those rather than a destination. The calendar became a date on Today. The wiki became the pages you reach by browsing Life. The inbox dissolved into two things that aren’t a room: a strip for the handful of items genuinely blocked without you, and a receipt trail for everything else.

The modes went internal. They’re still the response contract underneath: recognition still decides which of the five a thought becomes, and the five surfaces still render, but you never choose one. The choice appears as a chip on the capture card, and one tap flips it.

Deleting a concept from the screen is easy. Here’s the part I didn’t expect: it costs you something, and you have to pay it back deliberately.

What had to be added

Auto-filing creates an anxiety that manual filing doesn’t. If I put a note in a folder, I know where it is. If a machine puts it somewhere clever, I have to trust a system I can’t see, and the closest analogue in this category died of exactly that. Not of filing things wrongly. Of filing them invisibly.

So the restructure was symmetric: remove rooms, and add a legibility layer.

• A receipt on every capture. What it decided, where it went, as a link you can tap to land in situ. Later, a reason too: what the system saw that made it choose.
• A digest line on Today, deep-linking into the full trail.
• Confidence-tiered filing. High confidence lands silently and shows up in the digest. Low confidence doesn’t guess. It waits in the strip and asks, and it says why it’s asking.
• One canonical capture card, rendering identically everywhere a capture appears. That’s an architectural requirement, not a styling preference: if the same thought looks like a different object in three places, the receipt stops being evidence.

The lesson

I’ve come to think of it as a conservation law. Every concept you take off the screen has to come back as evidence, or the user just experiences your cleverness as loss of control. “Simpler” is not the same as “less visible”, and the second one, done without the first, is how a helpful system starts to feel like a system that’s up to something.