onBoulevard
A curated index of Zimbabwe’s built environment, designed around the day its developer leaves.
- Role
- Solo full-stack developer, freelance — from a designer’s Google Stitch files to handover
- Timeline
- ~6 weeks, handed over Sept 2026
- Stack
- Next.js 16 · React 19 · TypeScript · Tailwind v4 · Sanity · Airtable · Vercel
The problem
Zimbabwe’s built environment has no single index. Its record is spread across individual firm websites, press mentions and institutional archives, and the client’s own research — 208 people, firms and projects — lived in an Airtable base that doubled as an outreach CRM, next to deal values and contact notes that must never be published.
The brief arrived as two sentences, weeks apart. One was a solution: “I want to be able to drag things around and see it change.” The other was the problem: the client had to change entries themselves, without me. A designer had delivered the screens in Google Stitch; my job was everything from those files to the day the developer relationship ends — and that day shaped the architecture more than the screens did.
The solution
Sanity, with form editing against a live preview instead of a layout canvas. An editor opens the real page in the Presentation tool, clicks a heading, and edits it in a form beside it — in context, with no way to move the layout the design system depends on. Behind it, one module is the only code that knows where content comes from, so swapping local JSON for GROQ touched one file and no page.
Publishing reaches the public site through a signed webhook that invalidates only the cache tags an edit affects — no rebuild, and no connection held open per visitor. The CRM stays in use: a repeatable sync patches an allowlist of fields from Airtable and seeds editorial ones once, so it never overwrites curation. Search, filters and pagination are plain GET forms and links that work with JavaScript off and produce URLs a reader can cite.
- 01Five categories — People, Firms, Projects, Places, Products — with derived views like “people at this firm” that nobody maintains by hand.
- 02Click-to-edit against a live preview of the real page, with the layout out of the editor’s reach.
- 03Publish to live with no build: a signed webhook revalidates only the pages an edit touches.
- 04A repeatable Airtable sync that preserves relationships and never overwrites editorial work.
- 05Search, column filters and pagination that work with JavaScript disabled and compose into shareable URLs.
- 06Empty gallery, feature and related bands hold their place with a skeleton instead of collapsing the page.
Why this stack
Sanity
Hosted content, a schema in code, and a Presentation tool that gives click-to-edit on the real page. Free up to 20 seats, so a one-to-three-editor handover carries no recurring bill in a currency the client does not earn in.
RejectedPayload — what I would pick for a site I kept running, but it needs a database and a live server nobody would tend, and its visual editor is an enterprise feature. Storyblok, at $99 a month for a second editor. Tina and Keystatic, which make every typo fix a commit and a green CI run.
A standalone Studio package
The Studio is its own pnpm workspace package on its own dev server. It updates on Sanity’s hosting, so a CMS bugfix never needs a site redeploy, and the app’s build compiles a static site rather than a CMS.
RejectedEmbedding it as a catch-all Next.js route — the default every tutorial takes — which pulls in a second React, slows the build, and needed an exception carved into the content boundary.
Tagged fetch + a signed publish webhook
Every GROQ query is cached indefinitely under tags. On publish, a webhook verifies its HMAC signature and calls revalidateTag for exactly the documents that changed and the pages that reference them.
RejectedSanity’s Live Content API, which holds a third-party connection open behind every page view; timed ISR, which is never quite fresh; and a redeploy on publish, which makes publishing depend on a build pipeline nobody will be watching.
Airtable REST → an allowlisted patch
The API keeps record ids, so relationships resolve exactly and re-runs are idempotent. An allowlist of columns per table keeps deal values and contact notes out of the public dataset, even after someone adds a column.
RejectedThe client’s plan, a one-time CSV import. CSV flattens linked records to display names, so every person-to-firm edge would have been rebuilt by guessing at names.
Native GET forms and <details>
Search, filters and pagination are forms and links rendered by Server Components. A native disclosure gives the filter panel keyboard support and correct semantics without shipping any JavaScript.
RejectedA client-side popover filtering in the browser, which does nothing with JavaScript off and produces no URL to share.
- How a request moves through it
- 1An editor opens the page in the Studio’s Presentation pane, clicks the heading they want to change, edits it in the form beside it, and publishes.
- 2Sanity POSTs a webhook to the site’s revalidate route, signed with a shared secret; unsigned or forged requests are turned away with a 401.
- 3The route maps the changed document to every cache tag that reads it — a firm’s rename also stales the people and projects that show its name.
- 4revalidateTag marks those tags stale: the next visitor is served the previous copy once while the fresh one is fetched behind them.
- 5Every visitor after that gets the updated page — still static, with no build run and no developer involved.
Decision log
Hurdles that changed the shape of the system, in the order they were hit.
The CMS decision was made to cost one file
No CMS had been chosen when I started building, and whichever one won would change where every string on the site came from. So on day one all content went behind a single module of eight accessors, and they were async while they were still reading local JSON — functions with nothing to await, which looks like over-engineering.
That was the point. A synchronous boundary would have hidden where content came from but not whether reading it suspends, and a Promise spreads to every caller. When Sanity landed three weeks later, the swap to GROQ changed one file and not a single page or component.
Turning down the page builder
The client asked to drag things around and watch them change. Underneath was the real requirement — edit without a developer — and those are not the same thing. The content is five fixed tables and one profile layout generated from three route shapes; there is nothing to compose, and the design system is a deliverable in its own right.
So the client edits in context, on the real page, through forms, and cannot move the layout. The editing model is what keeps the design system enforceable after handover — reserved weights, square corners, one palette — rather than a limitation of the tool. Layout control was declined on the record, not left out by accident.
The Studio stands alone
Sanity’s setup offers to embed the Studio as a Next.js route, and the plan I had written said to accept. That plan had already carved an exception into the project’s most important rule — nothing in the app reads content except through the boundary — just to make room for a CMS.
Moving the Studio into its own workspace package removed the exception instead of documenting it. The two applications share a dataset, not code: each resolves its own React, the Studio updates without touching the site, and the schema still changes in the same commit as the types that depend on it.
Airtable owns the row, Sanity owns the profile
The client asked to export a CSV and load it once. But the CRM would keep being used after launch, so the import had to be repeatable — and the moment it is repeatable, two systems hold the same records. Overwriting on every run would quietly delete any biography written in the Studio.
The rule fits in one sentence for a handover: Airtable owns the index row, Sanity owns the profile. The sync patches allowlisted fields, seeds editorial ones once, and never deletes — vanished rows are reported, not removed. Field groups in the Studio say which fields the next sync will replace, because ownership shown in the interface beats ownership explained in a document.
Search spent the static premise, on purpose
Search had to survive a change of category — search People, click Firms, keep the query. A separate /search page would have kept the directory statically rendered, but it has no category tabs, so it could not meet the requirement; it only moved the dynamic rendering somewhere else.
So the directory reads its query string and renders on demand — the first time a public route stopped being static. The cost is bounded: query responses stay in the data cache, the home and detail pages stay static, and filtered views are noindex. I wrote the price down instead of absorbing it: “nothing to keep alive” is now slightly less true.
Skeletons instead of missing sections
Early on, most profile pages were a name and two paragraphs, and the rule was to omit empty sections. Each omission was honest, but together they made page length track how much the CMS happened to hold — and someone comparing practices reads length as a signal about the practice.
Gallery, Features and Related now hold their place with a skeleton. Borrowing an image as a video poster or showing a filler count was refused, because those assert something untrue; a grey block asserts nothing. Prose sections still drop when empty, because a paragraph-shaped placeholder is a claim about how much there is to say.
Trade-offs
No keystroke-level live preview
Rejecting the Live Content API cost instant refresh inside the preview pane. Editors keep click-to-edit, and a publish can show the old copy for one refresh while the new one regenerates — defensible, and easy to mistake for a bug.
The directory renders on demand
Search made the five category routes dynamic. The data cache keeps that to a render per request rather than a query per view, but a site promised to need nothing kept alive now needs slightly more than nothing.
Filtering happens in memory
Two filter fields are derived in the query projection, so matching runs over the rendered shape — one implementation for every category. It is linear, and past a few thousand entries in one category those fields have to become stored ones first.
What I took from it
The worst bugs were found by looking at output, not code. The sync resolved references before its first pass was queryable, so the relationship graph — the thing the product is positioned on — came out almost empty with no error, and healed itself on the next run. A build quietly rendered 18 pages of a 225-page site because the sync invalidated no cache. Both were fixed before handover; neither would have shown up in a diff.
A handover is a different brief from a build. Every decision here was weighed against the day nobody is maintaining it: no server to tend, no pipeline a publish depends on, and each silent failure mode written down where the next person will look first.
Never state something untrue about the data. The rule turned a hardcoded “1–10 of 118” into a real count, refused a borrowed video poster, and produced its corollary in the filter bar: a control must not hide its own output.