You're building a unified guest portal for a community-owned conservation lodge on the edge of Etosha National Park, on the wrong side of a 10Mbps satellite link.
The discipline skills: frontend architecture as a trade-off log with rendering strategy chosen per surface, a quantified performance envelope enforced in CI, accessibility designed into component primitives, a four-tier verification architecture with mutation testing, and SLOs with error budgets encoded as code alongside the application.
The AI-direction lesson: you direct AI as an architect now. You write the decision records, set the performance envelope, and specify the verification tiers before any prompt is written. Four areas reach the top of their range at the same time — frontend, performance, testing, observability — and they have to compose into one coherent system. AI will push toward more: more JavaScript, more sophisticated patterns, more optimisation, more alerts. The hardest judgment here is knowing when to stop, and being able to defend the answer.
Your Role
You own the whole architecture. No starter codebase, no PRD, no decomposition plan. The client describes three disconnected systems and a hard constraint; everything else is your design.
You decide what gets built and how. AI does the building — sometimes Claude Code, sometimes Codex CLI running through Claude Code as an MCP server. Every piece AI produces gets checked against the architectural invariants you wrote down first. The client is not technical, but she will have to defend every major decision to a conservancy board. If you cannot explain your choices in terms she can repeat, the work is not done.
What's New
Last time you rebuilt a restaurant group's ordering system into a real-time application and then made it production-ready with Terraform, security scanning, and your first AI infrastructure — skills, hooks, a custom agent, Codex CLI.
Three things change.
No codebase to inherit. P11 gave you a starter that encoded prior decisions. This time the three existing systems are described, not delivered. You write the architecture from a meeting transcript.
Four disciplines culminate at once. Frontend, performance, testing, and observability all reach the top of their range in this project. Each gets its own unit, but the units have to compose — one performance envelope, one component model, one verification tier design, one SLO set.
The hard part is restraint. AI generates sophisticated patterns on demand: micro-frontends, state machines, mutation testing for everything, alerts on every threshold. Most of them do not fit this problem. The judgment this project builds is knowing which ones do, and why.
Tools
- Next.js (App Router, server components, streaming), React 19, TypeScript — continuing.
- Playwright, Vitest, axe-core — continuing.
- Lighthouse CI, WebPageTest — extended with satellite-internet throttling profiles.
- k6 — load testing with bandwidth throttling.
- Stryker Mutator — new. Mutation testing for the critical-path suites.
- Sloth-style SLO generator — new. Observability-as-code.
- Grafana + Prometheus + Loki + Tempo, OpenTelemetry, Terraform — continuing, extended with the Grafana provider.
- Claude Code running as MCP server, Codex CLI as MCP client — new pairing.
- VS Code, Git + GitHub — continuing.
Materials
- Meeting transcript — how the project starts. Five measured paragraphs and one question.
- Business brief — the lodge's current public story, in the client's own words.
- Reference systems description — the three existing systems in domain terms, not code.
- Ministry wildlife format spec — a regulatory reporting schema. Discovered through conversation, not handed over.
- Tour operator API spec — a booking API the platform must accept. Also discovered.
- CLAUDE.md skeleton — a thin-brief template with verification targets left blank. You populate it from the decisions you make.
- Empty ADR directory — you write architecture decision records 0001 through 0007 across the project.
- Reference configs — Lighthouse CI, Stryker, Sloth SLO spec, Grafana Terraform, runbook template, handover template. Starting points with deliberate weaknesses you correct.