On Friday evening, a tab bar in one of our apps showed a small vertical scrollbar on the right. One pixel too many, caused by the focus ring. A detail. But the same pattern showed up in the settings navigation, in a row of short videos, in a date picker and in the table of contents of the reading pages.
The fix was made once, in the kit (krizaka-ui#295), with tests that failed before. Then every application received it by upgrading. That is the idea of a shared kit.
One kit, three platforms
Krizaka UI is the component library of all our sites and applications. It targets three platforms: the web (React), mobile (React Native with Expo) and the Mac (a Tauri window that shows the web components as they are, with a few desktop shapes).
On 17 October, the platforms page counts 156 components on the web, 79 on mobile and 112 on the Mac. The day before, it was 151, 76 and 107. Things move fast, and that is intended.
The rule that matters most is not a component, it is a decision: for each platform, every component must say whether it exists there, whether it is still to do (with a ticket), or whether it makes no sense there (and what to use instead). A new component without that decision fails the tests (krizaka-ui#283). On mobile, 22 gaps remain. I would rather see them written down than discover them in an app.
Reusing a component does not mean copying its code from one platform to another. It means keeping the same interface: the same name, the same variants, the same colours taken from the same design tokens. On the web:
import { Button } from "@krizaka/ui/button";
<Button variant="primary">Continuer</Button>;And on mobile:
import { Button } from "@krizaka/ui/native";
<Button variant="primary" label="Continuer" />;You can see the small difference: on mobile, the text goes through label. It is not perfect, and I say so. But someone who knows the web button already knows how to use the mobile one.
What it brings. One fix, made once. On the krizaka.com home page, the cards in a row did not line up their links when the sentences had different lengths. We fixed the card in the kit (krizaka-ui#187), with a test that measures the row in a real browser. Every card grid on every site benefited.
What it costs. The kit has a price too, and I prefer to say so. It has to be published, then every application has to upgrade. @krizaka/ui has had 48 published versions since it was created on 8 October. On 17 October, 22 package.json files in our repositories depended on it, and 8 were at least one minor version behind. A fix in the kit is not yet a fix in the app.
Native parity. 22 gaps means 22 places where a mobile app has to wait, or do something else. And there are contract differences, such as the button's label, that we will pay for sooner or later.
Count first, judge later
On 20 October, I showed Krizaka UI to developers around me. Their verdict came quickly: “It looks too AI. It is like every project generated by Claude: the same template, the same way of doing things, the same result at first glance.”
It stung, because it was true. The kit had been written with an assistant, and it bore its signature. Each detail, taken alone, could be defended. Together, they made a set we had seen a hundred times.
Earlier that same day, I had already had three things removed that bothered me: the buttons with cut corners, the blue and the hovers. Those three touches came from my own eye. The developers' verdict came after, and it said something else: the problem was not there. Before touching anything else, I asked for an audit, measured in a real browser on the live sites (krizaka.com, ui.krizaka.com, and the sites of two products).
How this was made
- Model
- Screen capture (no model)
- Engine
- not recorded
- Machine
- hosted by the provider (not our machine)
- Through
- Playwright capture of the live page
- Time
- not recorded
- Cost
- not recorded
- Sovkora run
- not recorded
- Note
- krizaka.com, desktop, dark, as audited on 2026-10-20 (before).
Here is what the audit found.
The fonts were Inter, Plus Jakarta Sans in weight 800 and JetBrains Mono: the default trio of Next.js templates. The krizaka.com hero was centred, with an uppercase mono label, a 64 px headline with an em-dash, a solid button and a ghost button. An uppercase mono label sat above 5 headings out of 6 on krizaka.com, and 6 out of 7 on Orazaka and on Yomizaka (the Yomizaka home page had 25). A band of 3 to 4 large numbers that count up followed every hero.
There were fake browser windows, with their three dots and their address bar, on ui.krizaka.com, Orazaka and Yomizaka. There were blurred elements: 11 on krizaka.com, 6 on the button page, 4 on the ui. home page, 3 on Yomizaka. On the button page alone, 231 elements had exactly the same 16 px radius.
The order of the sections was that of every SaaS page: label, headline, introduction, card grid, call to action, FAQ. The icons had the rounded 2 px stroke you see everywhere. And the brand's birds were decoration, when they should have been structuring something.
One remark I find fair: the words themselves were precise, and they stayed. The problem was not the text. It was the styling.
Rather than correct at random, we drew three complete directions on the same real screens (the krizaka.com home, the ui. button page, a Yomizaka lease check), in light and dark, on desktop and on phone.
How this was made
- Model
- Screen capture (no model)
- Engine
- not recorded
- Machine
- hosted by the provider (not our machine)
- Through
- Playwright capture of the live page
- Time
- not recorded
- Cost
- not recorded
- Sovkora run
- not recorded
- Note
- Static prototype of direction 1, Cahier, on the krizaka.com home (light, 1440 px).
How this was made
- Model
- Screen capture (no model)
- Engine
- not recorded
- Machine
- hosted by the provider (not our machine)
- Through
- Playwright capture of the live page
- Time
- not recorded
- Cost
- not recorded
- Sovkora run
- not recorded
- Note
- Static prototype of direction 2, Habitat, on the krizaka.com home (light, 1440 px).
How this was made
- Model
- Screen capture (no model)
- Engine
- not recorded
- Machine
- hosted by the provider (not our machine)
- Through
- Playwright capture of the live page
- Time
- not recorded
- Cost
- not recorded
- Sovkora run
- not recorded
- Note
- Static prototype of direction 3, Affiche, on the krizaka.com home (light, 1440 px).
Cahier: the French school-notebook grid (Seyès), a red margin where section numbers are written, the birds drawn like naturalist plates with their promise as a caption. Blue ink on paper in light mode, chalk on slate in dark mode.
Habitat: Montréal's modular architecture, slabs with a solid face, no blur, products stacked.
Affiche: a wall of posters, flat blocks of colour, headings that change width, the birds as stickers.
I chose Cahier. The reason fits in one sentence: it is the only direction where the birds organise the system instead of decorating it. They become the plates and the captions, and a generic prompt does not produce that on its own. Habitat was the safest, but the one most likely to slide back into a plain developer tool. Affiche was the loudest, hard to keep calm on a professional screen. We kept a stamp from it for the big moments.
The cut corners are gone. On every primary action button we had a corner cut at 45 degrees. A trick that gave the most important command on a page the look of a video game. A single 8 px radius, that of an input field, replaces it. Then, with Cahier, 2 px on controls and 0 on cards.
The blue. The old one, #196df5, is a saturated electric blue: the blue of almost every AI product. We drew five candidates as swatches, then on the real pages. The blue we kept is a steel blue, #3e76a8 in dark mode and #2a5e8d in light mode. Its hue (208) is exactly the complement of Orazaka's orange, which gives the advantage of a rule rather than a taste. The contrast of white text on it is 4.8 to 1 in dark mode and 6.7 to 1 in light mode, tested.
How this was made
- Model
- Screen capture (no model)
- Engine
- not recorded
- Machine
- hosted by the provider (not our machine)
- Through
- Playwright capture of the live page
- Time
- not recorded
- Cost
- not recorded
- Sovkora run
- not recorded
- Note
- krizaka.com, before and after the steel blue and the single button shape: desktop and phone, dark and light. The download figure is read live.
No more light following the mouse, no more card lifting. On hover, a card used to rise 3 to 4 px with a coloured shadow, and a glow followed the pointer. Hover now changes a colour, never a position: the card's background steps up one level and its rule darkens, in 150 ms.
No more blur. Shadows are a solid offset, like a sheet laid slightly to the side, never a halo. The bar is no longer frosted glass. On krizaka.com there are no blurred elements and no uppercase labels, against 11 blurred elements at the start; on ui.krizaka.com, one blur. Some pills and a few 12 and 14 px radii remain on krizaka.com: it is not finished.
How this was made
- Model
- Screen capture (no model)
- Engine
- not recorded
- Machine
- hosted by the provider (not our machine)
- Through
- Playwright capture of the live page
- Time
- not recorded
- Cost
- not recorded
- Sovkora run
- not recorded
- Note
- ui.krizaka.com home before the calm home, 1440 px, light.
How this was made
- Model
- Screen capture (no model)
- Engine
- not recorded
- Machine
- hosted by the provider (not our machine)
- Through
- Playwright capture of the live page
- Time
- not recorded
- Cost
- not recorded
- Sovkora run
- not recorded
- Note
- ui.krizaka.com home after the calm home, 1440 px, light.
The bug that taught us to test
On 20 October, I realised that the engine flow was not moving, neither on krizaka.com nor in the lab. The points of light were there, motionless.
The cause, once found, was tiny. The animation was sliding a stroke offset toward a value written without a unit, a bare number, while the other keyframe was a length. The browser cannot go from one to the other, so it did not animate. Each point jumped between two positions that drew it in the same place. A detail of one line of CSS (krizaka-ui#394).
What taught me the most is why nobody had seen it. All our checks in a real browser ran with reduced motion requested. They saw a perfectly calm page, as planned, and concluded that everything was fine.
Since then, motion is tested where it moves. Each animated element has two checks in a real browser: one with reduced motion off, which reads the animated property for about seven tenths of a second and requires different, finite values, and one with reduced motion on, which requires everything to stay at rest. And a lint rule: never an animated length without its unit.
The engine flow became EngineFlow in Krizaka UI, redrawn in SVG and CSS. The conduits are paths, the photons are zero-length points whose offset slides. We get, for free, what a canvas does not give: pausing off screen, text a screen reader can read, real buttons for the nodes, colour that comes from the tokens in both themes, and no layout shift. It weighs 2.2 KB compressed. On a phone, it stacks into a column.
How this was made
- Model
- Screen recording (no model)
- Engine
- not recorded
- Machine
- hosted by the provider (not our machine)
- Through
- The Krizaka Motion lab, played live in a browser
- Time
- not recorded
- Cost
- no cash cost (own hardware; electricity not metered)
- Sovkora run
- not recorded
- Note
- The engine flow page of the lab, played live; recorded from the lab's own pages. Muted.
One rule dominates everything else. Someone who asks for less motion in their system does not get an empty page: they get the resting picture, composed, with everything in place at once. And motion is code like any other: it can break without a sound. A motionless point of light hurt nobody, but it stayed invisible for as long as our tests looked elsewhere.
Why Orazaka has a Mac app
“Why a Mac app? The product already runs in a browser.” That is the right question, and the answer lies in Orazaka's promise.
Orazaka promises that your documents do not leave your premises. Yet those documents are files, in folders, on a machine. The more I looked at what that promise demands, the more a browser window looked like the wrong place to keep it.
A site can ask the user for a file, one at a time. For a tool to read a folder, watch it and write next to the PDFs that arrive, something else is needed: a folder chosen once, a permission that lasts, a limit the program enforces itself.
The app is called Orazaka Desktop. Its README describes it as version 0.1, a test build, for macOS 14 and later on Apple silicon. It runs Orazaka on your Mac or connects you to a hosted Orazaka, and it gives access to the folders you choose, and only those. The folder choice goes through the system's native picker, and you can revoke one folder, or all.
The limit lives in the Rust code (scope.rs). The app reads a path only if its canonical form is inside a granted folder. A .. or a symbolic link that leaves the folder is refused. And some folders can never be granted: your home folder itself, the root, the volumes, the system folders, ~/.ssh, ~/.aws.
The web part of the window has no file system permission, and the interface cannot launch a shell: it names a command from a short list, and the Rust side turns it into fixed arguments. I told the other side of this in the post on the agent on the user's machine.
I need to correct something. When I said “native app”, I should have said “Mac app”. It is not written in SwiftUI. It is Tauri 2: a Rust program that shows the interface in the system's web view (WKWebView on macOS). The decision is written in ADR-002 of Krizaka UI, with what we refused. Electron: a full Chromium to ship with each app, its own update system, one more attack surface. A native interface in SwiftUI or AppKit: a third implementation of every component, to keep in step with the web and the phone.
So here is the real answer to “why not a site”: not to draw our buttons differently, but for everything that touches the machine. The buttons are the web's. Everything specific to the window (the native menu, the menu-bar icon, notifications, file dialogs) belongs to the app, through Tauri, never to the components.
The same interface, not the same layout
Orazaka has a web client, a phone app and a Mac app. If we had to draw every screen three times, we would never finish, and after a month the three would no longer look alike. Here is how the code is arranged so that this does not happen.
The tokens. Colours and themes come from @krizaka/tokens. The package has a CSS output for the web and a native entry for React Native. A colour is changed in one place.
The components. @krizaka/ui for the web, and the same window for the Mac app. @krizaka/ui/native for the phone. The names are the same (Button, Badge, Card, Header, TextField), and so are the variants. The mobile app makes about forty imports from @krizaka/ui/native, and its local interface folder now holds only four small files, 177 lines in all.
The contracts. @krizaka/orazaka-shared holds the TypeScript types and Zod schemas of what the Orazaka API exchanges. The web client and the mobile app both depend on it. If the API changes, both change at the same moment, and the compiler says so.
And when the kit took on the “Cahier” identity (badges and buttons with sharper corners), tests in Orazaka's library broke: they checked a radius value. We fixed them so that they check the variant, not the radius (#54). A test that pins a visual detail breaks every time the kit moves, and that is not what we want it to watch.
The mobile app is an Expo project, in React Native. Android and iOS come from the same code. I am not claiming more than that: I have not checked any store publication.
Same interface does not mean same screen. On the web, an application's navigation is an icon rail on the left. On a phone, that rail becomes a tab bar at the bottom (the kit's changelog says so for RailNav). On the Mac, it is the web form inside the window, and the kit provides a title bar and a desktop density for it.
What is shared is the contract: names, states, tokens, words. What changes is the shape the same element takes on each machine.
Where we are going: the aim is that the same interface contract holds everywhere, web, Mac, Android, iOS. One component, one name, variants, states and tokens, and the shape each machine gives it. I do not promise a date, and I do not promise full coverage. If this did not work, I would see it in one thing: a screen that needs a behaviour specific to one machine, which the kit cannot express. Then the thing has to be added to the kit, not copied into the app.
