Northwind Coffee is finished: a catalogue people can filter, brew guides, shops, a contact form, chrome that editors own, two languages, copy that differs by reader, a governed paragraph, a staging environment and a cross-sell rail. Twenty-four Experience nodes, one route, no server.
The site was never the point. This is what it was arguing.
The argument
The second half of this series added generated types, contextual values, a second language, audiences, an editorial workflow, a staging environment and Adapters.
The content model barely changed. No new Contract was needed for localization or personalization. No Template changed for either. No renderer changed for either.
That is not because those features are shallow. It is because of two decisions taken in part 4, when both cost nothing and meant nothing: name is localizable, description is personalizable. Eleven parts of doing absolutely nothing, and then they were simply already right.
The alternative is not that those features are harder later. It is that they arrive as migrations — every field, every renderer, all at once, on a live site, usually under a deadline that is the reason anyone noticed.
The expensive part of a content model is never building it. It is the shape you did not leave room for.
The rules worth keeping
Six things that were doing most of the work:
Name a Contract for what it holds, never for where you first showed it. statement, not hero. This is what let one Contract have two Templates in part 3 with no migration and nothing copied — and it is what stops you needing hero-2 next year, when the second design arrives and the good name is spent.
Decide the localizable and personalizable axes when you create the field. You know whether a value can legitimately differ by language long before you know whether it will. producer is not localizable because a farm’s name is a proper noun; that is a question you can answer on day one.
Inline or reference is about independent life, not storage. Is it shared, is it separately publishable, is it worth finding in the library? A photograph is three noes. An origin is three yeses. And "somebody needs to review it" turns out to be a form of yes, which part 17 found the hard way.
A URL is a fact the CMS owns. Store identity and resolve the path; never build a URL out of a name. That one rule is why moving a page in this site breaks nothing at all.
Cardinality is a modifier. Enumerable composes with localizable, which composes with personalizable. Three modifiers wrap, rather than eight combinations being built.
Constrain the fields. allowedContractIds, allowedTemplateIds, allowedModes, allowedKinds — the difference between a model somebody can use confidently and one that needs a person standing next to it. And check them: an empty constraint is legal and means "any", as part 19 found out six parts late.
It is worth seeing the finished route table, because it never changed:
export default function App() {
const locale = useLocale()
const { header, footer } = useSiteChrome(locale)
return (
<div className="flex min-h-svh flex-col">
<Header content={header} />
<div className="flex-1">
<Routes>
<Route path="*" element={<ContentPage />} />
</Routes>
</div>
<Footer content={footer} />
</div>
)
}
Twenty steps, a catalogue, two languages, an approval flow, a staging environment — and one route, exactly as it was in part 2. Every page on the site arrived without touching it.
What Northwind deliberately does not do
Being clear about the edges is more useful than pretending there are none.
It does not render on the server. Every page is client-rendered from a static bundle, which is fast, cheap and nearly unbreakable — and means a crawler that does not run JavaScript sees the shell. Fine for this site; a real consideration for one that lives on search traffic — which is why the same site also exists with a server in front of it, in a companion series that starts where this one ends.
It does not author content by script. Content arrives by hand or by import, which is what the series teaches. The Content Management API exists — Pro and above — and content-as-code on it is a real thing to build; it is simply not what this sample demonstrates. Part 7 says why.
It is not a design system. The styling is deliberately plain, so that what you are reading is the content model rather than somebody’s CSS.
It is one site. Multi-brand, multi-site and shared libraries across them are real things this platform does and this sample does not exercise.
Where to go
The whole sample is at ebitex-io/samples — twenty tags, five checkpoint bundles, and a README that says what a bundle does and does not carry. The finished bundle is the fastest way to look around a complete content model without building one.
If you would rather start from your own thing than from Northwind: take the six rules above, model the first three Contracts of whatever you are actually building, and notice how much of the argument is about decisions you make in the first hour.
That is the honest summary of nineteen parts. Most of what a content model costs you later was decided very early, cheaply, by somebody who did not know they were deciding it.
