Multi-agent orchestration for building production apps with Claude Code

---

You describe what you want to build. Shipwright orchestrates 15 specialized AI agents through a 12-phase pipeline, from requirements to deployed and QA-tested.

Prerequisites

• Claude Code (Max plan recommended for higher rate limits)
• acpx - deterministic graph execution engine

npm ci -g acpx@latest

Quick Start

gh repo create my-app --template aster2709/shipwright --clone --private
cd my-app
npm ci
npm start

That's it. The CLI walks you through everything: pick build or feature, describe what you want, and the pipeline runs.

Alternative: Claude Code Skills (flexible, human-in-the-loop)

claude --dangerously-skip-permissions
# Then: /build a SaaS invoice platform with Stripe integration

How It Works

Q&A (5 rounds)  →  PRD  →  Research  →  Architecture  →  Design  →  Skeleton
 [interactive]    [review]              [review]

Implementation Plan  →  Backend + Frontend (parallel)  →  Testing
    [review]

Review  →  Audit  →  Deploy  →  QA  →  Learnings

Three interactive review gates where each document is rendered in your terminal and you can approve or suggest changes. The agent revises and re-presents until you're satisfied.

Dynamic requirements gathering with 5 rounds of adaptive questions. Each question is generated by the LLM based on your previous answers, covering users, auth, data model, deployment, and design.

Parallel implementation with separate backend and frontend sessions running concurrently via fork/join.

Shipwright pairs a deterministic graph engine (acpx fork) with Claude Code as the agent runtime. The graph handles scheduling, retries, and branching. The LLM handles reasoning and coding.

Agents (15)

Agent	Role
team-lead	Orchestrates pipeline, enforces completion, never idles
requirements-analyst	Clarifying questions, PRD with budget/deployment constraints
researcher	Best practices, cost analysis, technology recommendations
architect	System design, API contracts, data model, cost estimates
ui-designer	Design system, shadcn/ui + Magic UI scaffolding via 21st.dev
skeleton-builder	Repo structure, configs, empty modules that compile
implementation-planner	Task breakdown with dependencies and file ownership
backend-implementer	API routes, database, integrations, business logic
frontend-implementer	Pages, components, styling per DESIGN.md spec
test-engineer	Unit, integration, component tests
reviewer	Security, performance, correctness review
auditor	PRD coverage verification, gap analysis
deployer	Platform deployment, env config, cost-aware
monitor	Post-deploy health checks, CI/CD verification
qa-tester	Deep code audit, traces flows end-to-end, finds what breaks

Execution Modes

Mode	Engine	Completion	Best for
npm start	Interactive CLI + acpx graph engine	Guaranteed	Recommended for all users
/build skill	LLM orchestrator (Agent Teams)	Best-effort	Exploratory work, flexibility
/feature skill	LLM orchestrator (Agent Teams)	Best-effort	Quick feature additions

Skills

Skill	When
/build	New app from scratch — full 12-phase pipeline
/feature	Add features to an existing codebase — 8 phases
/audit	Gap check against PRD anytime

Agent Communication

1. Direct messaging — teammates talk to each other for questions and handoffs
2. Shared task list — work items with dependencies and status tracking
3. docs/ folder — formal outputs that become the project's source of truth

The team lead is the heartbeat — actively checks progress, spawns next phases immediately, never waits passively.

Documentation

Agents produce these artifacts in docs/:

File	Producer	Contents
PRD.md	requirements-analyst	Product requirements, user stories, constraints
RESEARCH.md	researcher	Technology recommendations with cost analysis
ARCHITECTURE.md	architect	System design, API contracts, deployment plan
DESIGN.md	ui-designer	Visual design system, component library choices
IMPLEMENTATION_PLAN.md	implementation-planner	Tasks with dependencies and file ownership
QA-REPORT.md	qa-tester	Browser test results with screenshots

Architecture

cmux / tmux                          (observe agents in split panes)
  └── acpx                           (deterministic graph engine — schedules phases)
       └── Claude Code               (agent runtime — executes each node)
            ├── 15 agent definitions  (each owns one phase)
            ├── docs/                 (inter-agent communication via artifacts)
            └── .claude/skills/learnings/  (reusable patterns from past builds)

Self-Improvement

After successful builds, reusable patterns are saved to .claude/skills/learnings/. Future builds reference these to avoid rediscovering the same solutions. The more you build, the better it gets.

Recommended

• cmux or tmux - split-pane agent visibility
• 21st.dev Magic components - UI design inspiration

License

MIT