The week the architecture was drawn

The beginning

The project was conceived on a Sunday — 21 March 2026 — and started the next day. By the end of that first week, eight numbered Implementation pages had landed: the three-project Claude system, the auto-memory protocol, the three-layer prompt architecture, the client onboarding pipeline, the two-Claudes topology, the NZ Privacy Act compliance posture, the skills-and-plugins concept, the early privacy filter. Eight foundations. Eight days.

The pace wasn't unusual for a brain that runs hot in concept work. The structural choice was that every foundation went down as a written page, not a head-note. The wiki started before the product did. Months later, the wiki is still the durable spine.

Capture before code. The wiki started before the product did.

Two Claudes

The signature architecture from Day 1 was a topology — not a feature. Two Claude instances, in different roles, communicating asynchronously through a shared Notion workspace. Jamie's Claude on Claude Max, running on the desktop, with read access to every client's workspace via local MCP. The client's Claude on the client's iPhone, on the free tier, with read and write to their own Notion only. The two never talked directly. The workspace was the bridge.

The split was load-bearing in two directions. Jamie's Claude could see across clients without violating the privacy boundary — it read each workspace through the client's own integration token, not through any shared backend. The client's Claude couldn't reach Jamie's workspace at all. Privacy was a property of the topology, not a permission.

The architecture has evolved since. The client's Claude is now a Custom Agent inside Notion rather than the standalone Claude app, and the Hybrid Member Model has replaced the guest model. But the core choice — two agents, one workspace, async through writes — survived every pivot.

Privacy is a property of the topology, not a permission.

Three layers

The first design that wasn't just a topology was the prompt architecture for the client's Claude. Three pages in Notion, each holding a different layer:

• Agent Profile — the skill layer. Who the client is, how to support them, what to do in specific situations.
• _Agent Protocol — the protocol layer. Workspace rules, confirmation requirements, voice command mappings, escalation boundaries, permission scope.
• _System Index — the memory layer. Current workspace state, recent changes, patterns.

The client's Claude reads a different combination of pages depending on the task. A simple "what's my next task?" needs Memory plus Skill, not Protocol. On a free Claude tier with smaller context windows, that mattered. Different layers, different update rates, different failure modes — each page is one thing, and the failure mode tells you which page to check.

The split was grounded in cognitive artifact theory — Norman's tools, protocols, and skills distinction, citation included in the page. Future iterations would refine the split, retire some pages, add others. The three-layer split itself was the foundation.

The prompt is architecture, not copy.

Built for compliance

By the end of Week 1, two pages had landed that named what the system would be regulated against. Implementation 08 set the NZ Privacy Act posture — a structural-privacy property (no Neuro Support backend, all client data on the client's own Notion plan), a two-flow client access model, cross-border processing clarity. Implementation 12 set the IF Funding catalogue — the funding mechanism the service would operate inside, with precedents documented for AI-powered support tools.

Neither was a marketing decision. Both were structural choices made before the product had its first prospect. The product would operate inside NZ disability funding from the start; it would never own the data plane. The privacy posture wasn't a policy bolted on later — it was the architecture itself.

Compliance is built in or it isn't there at all.

What's next

Week 1 set down the architecture. Week 2 brought the brand system, the early concept library, and the first naming decisions — the Seven Principles took their final form, Neuro Support stabilised as the name, the Protocol document was drafted.

Eight days in. A long way to go.