Claude Starter Kit

Autonomous AI-driven development agents for Claude Code. One command to scaffold, then let Claude build your TypeScript project.

Framework-agnostic: Supports Angular, React, Vue, Svelte via frontend profiles. Optimized for Nx monorepo but works with standalone projects too.

Quick Start

npx create-claude-workspace

This scaffolds the .claude/ folder (agents, templates, profiles) into your project.

Then initialize your project:

claude --agent project-initializer

Answer the discovery questions (non-technical). Claude generates the full pipeline:

• CLAUDE.md — architecture & conventions
• PRODUCT.md — product strategy & monetization (via product-owner agent)
• TODO.md — phased development plan (via technical-planner agent)
• MEMORY.md — progress tracking

Start Autonomous Development

npx create-claude-workspace run                          # start (1 agent)
npx create-claude-workspace run --concurrency 3          # 3 parallel agents
npx create-claude-workspace run -i                       # interactive mode (user input only)
npx create-claude-workspace run --task-mode local        # force local task mode
npx create-claude-workspace run --resume                 # resume from saved state

Or interactively: claude --agent orchestrator, then /ralph-loop:ralph-loop Continue autonomous development according to CLAUDE.md

Scheduler v2 (Multi-Agent)

The scheduler replaces the single-orchestrator model with parallel agent execution:

• TS scheduler handles workflow (task picking, git, CI, merges) — zero AI tokens on routing
• Claude agents get short, focused prompts (~500 tokens vs ~20k+ in v1)
• Each task = fresh context — no accumulating context across tasks
• N agents in parallel — independent tasks run concurrently in separate git worktrees

npx create-claude-workspace run                          # start (1 agent)
npx create-claude-workspace run --concurrency 3          # 3 parallel agents
npx create-claude-workspace run -i                       # interactive mode
npx create-claude-workspace run --task-mode local        # force local mode
npx create-claude-workspace run --help                   # all options

Task modes

The scheduler supports three task modes (--task-mode):

Mode	Source	Use case
local	TODO.md / tasks.json	Solo development, no issue tracker
platform	GitHub/GitLab issues	Team workflows with issue tracking
interactive	User TUI input only	Ad-hoc tasks, exploration, prototyping

Auto-detected at startup: if the repo has a remote with status:: labeled issues → platform, otherwise → local. Override with --task-mode flag or set Task Platform in CLAUDE.md via project-initializer.

The --interactive (-i) flag implies --task-mode interactive. Type task descriptions directly in the TUI — each becomes a task that goes through the full pipeline (plan → implement → test → review → commit). Pending tasks are shown in the status bar queue.

Communicate at runtime

Write to .claude/scheduler/inbox.json while the scheduler is running:

[
  { "type": "add-task", "title": "Add dark mode toggle", "taskType": "frontend", "complexity": "S" },
  { "type": "message", "text": "Focus on performance optimization" },
  { "type": "stop" }
]

The scheduler reads the inbox every iteration and processes messages immediately.

Proactive problem resolution

Agents never dismiss problems as "pre-existing." The scheduler injects a report_issue MCP tool into every agent. When an agent discovers an out-of-scope problem:

• Non-blocker: calls report_issue(severity: 'non-blocker') and continues working
• Blocker: calls report_issue(severity: 'blocker'), wraps up, and returns

The scheduler records the issue locally (discovered.ndjson) and creates a platform issue (GitHub/GitLab) if available. Discovered issues enter the task queue on the next iteration.

npx Options

npx create-claude-workspace [directory]    # scaffold into a specific directory
npx create-claude-workspace --update       # overwrite agents with latest version
npx create-claude-workspace validate       # check prerequisites (Claude CLI, auth, git)

Agents

Agent	Purpose
project-initializer	Orchestrates full setup pipeline
product-owner	Product strategy, features, monetization, prioritization
technical-planner	Translates PRODUCT.md into phased TODO.md (with quality validation)
devops-integrator	GitLab/GitHub: milestones, issues, branches, MRs/PRs, post-merge cleanup
deployment-engineer	CI/CD pipelines, deploys, rollbacks, smoke tests, DB migrations
orchestrator	Development cycle: 10-step workflow with error recovery
ui-engineer	Frontend: components, styling, state, a11y (any framework)
backend-ts-architect	Backend: API, DB, business logic, DI, services
senior-code-reviewer	Code review: security, performance, architecture, duplication
test-engineer	Write and run tests: Vitest/Jest auto-detect + Playwright E2E

Pipeline

project-initializer
  |-> Discovery Conversation   -> CLAUDE.md
  |-> product-owner            -> PRODUCT.md (features, monetization, MVP)
  |-> technical-planner        -> TODO.md (phases, tasks, dependencies)
  |-> devops-integrator (opt)  -> GitLab/GitHub milestones + issues
  |-> deployment-engineer (opt) -> CI/CD config + ## Deployment in CLAUDE.md
  |-> MEMORY.md

ralph-loop -> orchestrator agent (autonomous development)
  |-> Health check (track iterations, validate agents, sync main, ingest issues)
  |-> Pick task from TODO.md (+ dependency check, phase re-evaluation)
  |-> Create branch (devops-integrator)
  |-> Plan (backend-ts-architect first for fullstack, then ui-engineer)
  |-> Implement (+ DB migrations via deployment-engineer)
  |-> Write tests (test-engineer: auto-detects Vitest/Jest + Playwright E2E)
  |-> Build, lint & test (nx affected or explicit projects)
  |-> Visual verification (Playwright, UI tasks only)
  |-> Review (senior-code-reviewer, with TESTING context)
  |-> Rework -> Re-review (max 5 cycles, then escalate/skip)
  |-> Commit (source + tracking) + push + MR/PR (devops-integrator)
  |-> Post-merge (solo: merge to main / team: continue on next branch)
  |-> Next task

Fully Unattended Setup (PLAN.md)

For server/CI environments where no human interaction is possible, create a PLAN.md file in the project root before starting the scheduler. The project-initializer reads it and skips the discovery conversation.

# 1. Copy the template
cp .claude/templates/PLAN.md PLAN.md

# 2. Fill in your project details, credentials, and preferences

# 3. Start — no interaction needed
npx create-claude-workspace scheduler

Required Plugins

Install before starting:

# Core (required for interactive mode / ralph-loop)
claude plugin add ralph-loop
claude plugin add @anthropic/plugin-playwright

# Recommended
claude plugin add frontend-design        # UI generation without Figma
claude plugin add claude-md-management   # CLAUDE.md maintenance
claude plugin add commit-commands        # git workflow shortcuts

# Git integration (optional, pick one)
# GitHub: install gh CLI + run: gh auth login
# GitLab: install glab CLI + run: glab auth login

# Optional MCP Servers
# Figma MCP — configure in Claude Code MCP settings with API token

Frontend Agents

Framework-specific agents are shipped in .claude/agents/. During project initialization, the correct agent is selected based on the detected framework.

Shipped agents:

• angular-engineer.md — Angular (signals, standalone, zoneless, SSR, TestBed, theme system)
• react-engineer.md — React (hooks, server components, Testing Library)
• vue-engineer.md — Vue (Composition API, Pinia, Vue Test Utils)
• svelte-engineer.md — Svelte (runes, $state, SvelteKit)

Only ui-engineer, senior-code-reviewer, and test-engineer read the frontend profile. Backend and infrastructure agents are unaffected — they don't receive framework-specific context.

Limitations

• Requires Claude Code with Agent tool support (subagent_type parameter)
• ralph-loop and other plugins are Claude Code community plugins — install via claude plugin add <name>
• Git integration requires gh (GitHub) or glab (GitLab) CLI to be installed and authenticated
• Shell commands assume bash (Unix syntax) — works via Git Bash or WSL on Windows
• Dev server cleanup uses npx kill-port — requires npm/npx to be available

What Gets Generated

File	Content
CLAUDE.md	Architecture, coding standards, DI, workflow mode (solo/team), autonomous workflow
PRODUCT.md	Vision, target users, features (MVP/v1.1/future), monetization, metrics
TODO.md	Phased tasks with type/dependencies/complexity/acceptance criteria, checkboxes ([ ] / [x] / [~])
MEMORY.md	Session state: current task/step, iterations, complexity tracking, done, next, decisions, blockers