How We Work with Claude
This section documents the AI-assisted development workflow behind the marketplace — not as a retrospective, but as a live system. Every template, decision, and domain reference is used in active Claude Code sessions.
The files in .claude/ at the project root are loaded automatically by Claude Code at the start of each session. CLAUDE.md sets the hard constraints; the prompts and ADRs fill in the context.
The three-layer system
Each Claude session draws from three layers of context:
- CLAUDE.md — the law. Rules that must never be broken: async-only routes, inline styles, Pydantic v2 syntax, no Tailwind classes.
- Prompt templates — the procedure. Structured task specs that constrain scope, list files to read, and define what "done" looks like.
- ADRs — the memory. Why we made the decisions we made, so Claude doesn't re-litigate settled questions or repeat past mistakes.
◧7 templates
Prompt Library
7 reusable prompt templates — project context, new features, new apps, debugging, refactoring, code review, and testing. Copy and fill in the brackets.
Read more →
◈5 ADRs
Architecture Decisions
5 ADRs documenting why we chose async FastAPI, JSONB storage, WebSocket polling, MDX docs, and importlib dispatch — with the exact gotchas we hit.
Read more →
⊟Filterable
Domain Glossary
Actuarial terms, formulas, and validation benchmarks: Gompertz mortality, GBM pricing, Solvency II, IFRS 17. Filterable by keyword.
Read more →