You're fixing a booking system for a river eco-tourism company in Luang Prabang, Laos — diagnosing stale data from broken state architecture, then building your first project memory file.
The discipline skills: client-side vs server-side state management, cache invalidation strategy, diagnosing stale data bugs where the database is correct but the page lies, and authoring CLAUDE.md and AGENTS.md as session-start infrastructure.
The AI-direction lesson: you move from using AI tools to configuring them. Last project had a pre-built CLAUDE.md — you saw how it changed AI's output from the first prompt. This project, there's no memory file. You notice the gap. Then you write one yourself. The difference between using someone else's infrastructure and building your own is the difference between benefiting from it and understanding what makes it work. Specific, testable constraints produce specific, verifiable output. Vague entries produce vague compliance. The memory file you write becomes infrastructure that shapes every future session.
Your Role
You fix a booking system that mostly works but disagrees with itself. The database updates correctly when bookings change. The pages keep showing the old data — the client-side cache doesn't know anything changed.
After fixing the state architecture, you build your first project memory file — writing down the stack, the patterns, and the constraints in a form that Claude reads at session start.
What's New
Last time you added CI/CD and connected AI to a database for a Singapore VFX studio. You worked with a codebase you didn't write, built infrastructure around existing code.
Two things change.
You fix broken state, not broken process. The CI/CD pipeline was about automating quality checks. This is different. The code runs. The database is correct. But the page lies. Finding why the page lies when the database tells the truth — that's the new terrain.
You build AI infrastructure yourself. Last project had a pre-built CLAUDE.md. You experienced how it changed AI's output from the first prompt. This project, there's no memory file. You notice the gap. Then you write one yourself. The difference between using someone else's infrastructure and building your own is the difference between benefiting from it and understanding why it works. AI dependence is a documented career hazard — developers who always accept AI's defaults lose the ability to evaluate them. Building your own infrastructure, rather than relying on what's handed to you, is how you maintain agency.
Tools
- Next.js, React, TypeScript, PostgreSQL, Prisma — continuing.
- PostgreSQL MCP server — continuing from P8. AI reads the database directly.
- CLAUDE.md / AGENTS.md — project memory files. New — you author these yourself.
- VS Code + Claude Code — continuing.
- Git + GitHub — continuing.
Materials
- Existing booking system — a Next.js + PostgreSQL application with the state management bug. Functional but unreliable when bookings change.
- Voice memo from Kham — how the project starts. He tells you the problem through a story about a couple from Melbourne.
- Backlog notes — Kham's own list of the pieces of work he needs done, kept as a reference; the units below sequence them.
- AGENTS.md reference — a brief guide to the cross-platform project memory standard.