The Shift: From "AI Assists" to "AI Architects"
In 2024, AI was a helper—you describe a feature and AI writes 50% of the code. In 2025, AI is the architect. You describe the domain, and AI generates tested, deployable modules.
Case Study: hda_collection
Our largest module, built entirely with AI assistance:
- 24 models — ticket categories, collection sessions, deposits, treasury, incentives, budgets
- 26+ views — forms, trees, dashboards, reports
- 180+ tests — security, accounting, state transitions
- Build time: 3 days (with vibe coding) vs. 20 days (traditional)
- Post-launch defects: 2 (both in edge cases, not boilerplate)
Our Process: CLAUDE.md → /audit → /implement
Phase 1: Document Everything
We maintain a CLAUDE.md file with:
- Module purpose and dependencies
- Odoo version gotchas (v18 → v19 breaking changes)
- Security rules and access control matrix
- Data model relationships (which tables reference which)
- Code patterns (Python ORM, OWL v2, QWeb templates)
This becomes the context for Claude Code. Instead of searching docs, Claude reads your project's CLAUDE.md.
Phase 2: Audit Existing Code
Run /audit to identify:
- Anti-patterns (hard-coded colors, layout thrashing)
- Performance issues (missing lazy loading, expensive animations)
- WCAG accessibility gaps
Score each dimension 0-4, then prioritize fixes.
Phase 3: Generate & Iterate
Tell Claude Code what you need:
"Create a Python model for revenue collection sessions with:
- Multi-point support (museum, amusement park, etc.)
- 26+ ticket category types
- Real-time collector POS integration
- Immutable accounting journal entries
- See CLAUDE.md §5.5 for schema details"
Claude reads the context, generates the model with computed fields, constraints, and tests—all at once.
What AI Does Well
- ✅ ORM boilerplate (model declarations, fields, constraints)
- ✅ CRUD views (form, tree, search, action definitions)
- ✅ Tests (TransactionCase, assertions, fixtures)
- ✅ Admin screens (simple CRUD interfaces)
- ✅ Documentation (docstrings, comments)
What Still Needs Humans
- ❌ Domain knowledge ("what does a 'cutting order' actually do?")
- ❌ UX decisions ("where should this button go?")
- ❌ Performance tuning (when to add indexes, optimize queries)
- ❌ Compliance edge cases (ZATCA e-invoicing rules, audit trails)
- ❌ Security architecture (who can see what data?)
The Future
By 2026, AI agents will read your requirements doc, your CLAUDE.md, and the Odoo source code—then generate entire modules from scratch, run tests, fix failures, and deploy without human intervention.
We're already close.