A subscription marketplace platform spans three repositories — a Next.js frontend, a Ruby on Rails backend, and a Ruby AI service — and wanted to ship features faster without diluting the architectural patterns the team already relied on. We configured Claude Code into a multi-agent development system: per-repo context files, parameterized skills for recurring scaffolding, and specialized coder and QA agents wired to PostgreSQL, Figma, and a headless browser through MCP servers. A human orchestrator drives the system end-to-end; the agents handle implementation, testing, and visual verification between checkpoints. The Products feature — a non-trivial extension covering catalog, variants, inventory, subscription options, AI-assisted content, and media management — served as the harness, rebuilt from a clean slate on every iteration to validate the setup.
Generic AI coding assistants treat every repository as new and every convention as optional. That doesn't scale across a three-repo platform with established patterns for service objects, authorization policies, request specs, API client hooks, and SQS-backed inter-service communication. The team had minimal existing AI tooling configuration, no multi-agent workflow, and no automated verification loop. The agents needed deep repository context, narrow responsibilities, and the discipline to keep iterating until tests passed and frontends matched designs — not a freeform copilot.
The system is a configured instance of Claude Code with four layers: repository context, parameterized skills, specialized sub-agents, and MCP-mediated tool access. Each layer carries a piece of the platform's grammar so that generated code lands inside existing patterns rather than next to them.
- A CLAUDE.md per repository plus a cross-repo root, encoding architecture, naming, directory layout, and test conventions
- Slash-command skills that scaffold the recurring shapes — Rails endpoints with controller, route, policy, serializer, and request spec; Next.js pages with hooks and TypeScript types
- Specialized coder agents for backend, frontend, and the AI service, each with a narrow context window and only the tools it needs
- QA agents that run RSpec on the backend and exercise the frontend in a headless browser against Figma frames, routing failures back to the coders
- MCP servers for PostgreSQL schema inspection, Figma, and browser automation, replacing manual asset export and screenshot comparison
- A clean-slate iteration loop — the feature rebuilds from scratch every cycle so setup gaps surface as failures, not as patches
Rebuilding the feature from a clean slate every cycle is what turns the setup into a deliverable.
The Products feature shipped to the platform's staging environment — catalog, product create/edit, variants, inventory tracking, per-product subscription options, SEO metadata, and AI-assisted content generation — built end-to-end through the agentic workflow with human review at the planning and merge-request checkpoints. The agentic system stayed in the repositories as committed configuration, so the team can drive subsequent features with the same orchestrator-and-agents model rather than treating Claude Code as a generic assistant.