Building the Control Centre that built itself

The brief

Engine Labs sells agentic systems that replace repeatable operational work. The positioning only works if the first thing a visitor touches behaves like the product — not a contact form, not a Calendly link, not a deck.

So this site had to ship with a working Control Centre: type a problem, get clarifying questions, get an Engine recommendation and a draft scope band. The marketing site and the demo are the same artefact.

Constraints I set for myself: one continuous build session, under 24 hours wall-clock, under A$50 cash outlay (API usage + hosting — not counting tools I already pay for monthly).

This post is the case study for this repo — the site you are reading.

What shipped

A full marketing surface on Next.js 16 (App Router), TypeScript, Tailwind CSS v4, and MDX for long-form Engine specs and Lab posts.

APIs (same sprint)

POST /api/briefs/chat — OpenRouter recommender (Claude Opus 4.7 by default), clarify/recommend JSON, per-browser session spend cap.
POST /api/briefs — Persists briefs and recommendation payloads locally.
POST /api/contact — Resend to hello@enginelabs.com.au with a local inbox backup.

The Control Centre is not a mock. Submit a real brief on the homepage or Control Centre and you get a scoped conversation ending in a structured recommendation — the same class of agent I would deploy for a client Sales or Ops Engine.

How it was built (honest order)

Contracts first. The MSA, Addendum and Pricing Schedule were already written. Every page line that sounds like a promise traces back to a clause — including what the Centre is allowed to decline.
Product shape. Eight Engines, six verticals, Lane C positioning (“stop hiring for repeatable work”). Strategy docs and the live site share one source of product facts so copy does not drift.
IA + copy. Page structure locked before build. Engine pages use shared spec sections so eight long breakdowns stay consistent without eight one-off layouts.
Control Centre. Recommender prompt written and wired to the API first, then the homepage widget and the dedicated Control Centre page, with deep-links that prefill a brief when you arrive from an Engine or vertical.
Polish pass. Developer-inspired palette, dark mode, motion with reduced-motion support, scroll progress, footer consolidation — all in the same push.

Most of the typing was me directing; most of the keystrokes were Cursor agents running against this monorepo. That is intentional — it is the same delivery model I sell.

Stack (and why)

Next.js 16 + MDX — static Engine pages, one dynamic recommender API, easy Vercel deploy
Tailwind v4 + shadcn — fast iteration on section bands without a design system committee
OpenRouter — model routing without locking to one vendor; Opus 4.7 for quality on scope conversations
Resend — transactional email for contact and brief handoff
Vercel — hobby-tier deploy ($0 at launch scale)
Local .data/ stores — briefs and usage counters on disk during dev; git-ignored

No Supabase, no Stripe, no authenticated client portal in this sprint — those are real Week-2+ product surfaces, not blockers for proving the Centre on a marketing site.

Time and money

Wall-clock build — ~22 hours (one session, a short break, then ship).
OpenRouter (dev + smoke tests) — ~A$35.
Vercel — A$0 on the hobby tier.
Domain / DNS — already owned.
Total cash for this ship — under A$50.

The A$50 cap forced discipline: no bespoke CMS, no multi-environment staging theatre, no animation rabbit holes. Every hour had to move a route from “missing” to “live.”

Decisions that mattered

Centre as hero, not contact form. If the pitch is agentic, the first interaction must be agentic. Everything else is supporting evidence.
Eight Engines, routed by the Centre. Broad menu, narrow conversation — the recommender picks fit, not the visitor.
Boundaries as copy. /what-we-dont-do, contract-aligned FAQs, and a decline path in the recommender. Senior buyers close faster when you show what you will not do.
MDX spec components. Engine pages are data-heavy; components prevent eight one-off layouts.
Dogfooding the build. This site is Founder Engine work productised as marketing — prototype-speed delivery with a documented handover (this Lab, the repo README, the recommender prompt in-tree).

What is still open

Being precise beats pretending everything is finished:

Methodology PDF downloads on /methodology are still placeholder cards — the descriptions are real; the files are not attached yet.
Draft SOW PDF export from the Centre — recommendation JSON exists; PDF generation does not.
Stripe AU + Documenso — payment and e-sign stay manual until the contract pack is wired.
Authenticated client mode — project board, tickets, handover pack — post-launch product, not launch blocker.

None of that invalidates the case study. The claim is: a credible AI build studio site with a live scoping agent, shipped in one day for the price of a nice dinner.

What I would do differently

Start from the recommender API on hour one, not hour twelve — UI without a working brain wastes time.
Check Turbopack + MDX plugin constraints earlier (serialisable next.config options caught me once).
Capture OpenRouter spend in the UI during dev so the A$50 cap is visible, not discovered in the dashboard afterwards.

Try it

Brief a real problem — yours, not the example chips — on the Control Centre. If the recommendation is wrong, that is useful data too; tell me at hello@enginelabs.com.au.

Tools: Next.js 16, TypeScript, Tailwind CSS v4, MDX, OpenRouter (Claude Opus 4.7), Resend, Vercel, Cursor.