You're the sole developer for a provincial agriculture extension office in Goroka, Papua New Guinea — building a digital data-collection and reporting platform from paper, ahead of a national digital-reporting deadline.
The discipline skills: offline-first PWA architecture (IndexedDB, Background Sync, an explicit conflict-resolution policy); a shared data contract behind multiple audience-shaped reports; partner-indicator translation against IFAD and World Bank definitions; English and Tok Pisin in the data-collection app, English-only in the partner exports; one instrumentation producing audience-shaped verification evidence; an ADR set, a threat model, and a handover packet sized for a small government IT team. The work is communication as the load-bearing deliverable.
The AI-direction lesson: every artifact has an audience, and AI's first draft is almost always written for the wrong one. The same architectural decision has to read one way to a national director, another way to a partner program officer, another way to a developer in a different province who has not met you. AI can draft each version; you decide which one the audience needs, and you hold the shared contract so the numbers reconcile. AI's audience-blindness shows up everywhere — generic SLOs for an offline system, Kubernetes proposals for an eighteen-person agency, indicator translations that match the format but use the wrong definition. Specify the constraints before AI generates anything; read every output for the audience-fit its defaults miss.
Your Role
You're on a focused engagement. Build the platform, document it well enough that the rollout to other provinces can proceed, and leave. Grace Koma's office has eighteen staff, twelve field officers, no in-house IT, and a national deadline. The deliverable is the running system plus the documents the next team reads first.
Your role is supervisory. AI does the implementation across the data-collection app, the sync backbone, the multi-audience reports, the public dashboard, the partner exports, and the threat-model and ADR drafts. You hold the architecture, calibrate each artifact to its audience, and review every output for audience fit. The AI infrastructure you designed from scratch last time carries forward as practice — rules and skills rewritten for this project, the shape known.
What's New
Last time you ran a performance-and-resilience engagement for a luxury pearl-jewelry e-commerce site in Bahrain — image architecture, circuit breakers, a two-tier checkout, AI infrastructure designed from scratch.
This project moves on three axes.
Communication artifacts are the deliverables. Every unit produces an audience-shaped artifact: the brief, the ADRs, the multi-audience reporting backbone, the impact dashboard, the partner-format export plus API documentation, the handover packet. The verification evidence is shaped four ways from one instrumentation.
You start from paper. No existing codebase. The materials zip gives you scanned forms, the partners' indicator definitions, reference docs, and empty AI-infrastructure starter files. Everything else you build.
The client speaks in formal memos and volunteers very little. Grace answers what she's asked. Offline-first, partner-format specificity, the rollout plan, multi-language, documentation-as-deliverable — all hidden until you ask. The brief you write from discovery is the first audience-shaped artifact; everything downstream is calibrated against it.
Tools
- The Goroka Extension Platform stack — built from scratch. Next.js 15 for the data-collection PWA and the public dashboard, an Express API for sync and reporting, PostgreSQL on managed PaaS, single-region deployment in Singapore or Sydney.
- IndexedDB, the Background Sync API, and a service worker pattern — new. The offline-first foundation. You design the conflict-resolution policy explicitly.
- next-intl — new. English and Tok Pisin in the data-collection app.
- OpenAPI / Swagger — continuing. Source of truth for the partner-export contract.
- OpenTelemetry, Grafana, Prometheus, Loki, Tempo — continuing. Audience-shaped dashboards from one instrumentation.
- Lighthouse, WebPageTest, Playwright, Stryker — continuing. Mobile profile, offline sync E2E, mutation testing on indicator translations.
- adr-tools — new. The ADR set under
docs/adr/is the canonical handover artifact. - Claude Code with project memory, path-scoped rules, skills, and hooks carried forward, rewritten for this project's surfaces.
- VS Code, Git, GitHub — continuing.
Materials
- A formal memo from Grace — the first-contact artifact. She names three audiences and one constraint (connectivity). Hidden constraints surface only through the discovery exchange.
- The office's current paper forms — scanned PDFs.
- The IFAD and World Bank indicator definitions — the basis for every indicator-translation decision.
- Reference documentation for offline-first PWA patterns, ADR practice, and threat-modelling.
- Empty AI-infrastructure starter files — a
CLAUDE.md,AGENTS.md, and empty rules, skills, and hooks scaffolds. You rewrite from last project's pattern; you do not redesign the architecture.
No PRD, system design, API contract, threat model, ADR set, handover packet, test suite, dashboards, deployment configuration, or language tables. Those are your deliverables.