name: turkey-build description: Multi-agent app builder with 7 modes - greenfield, iteration, bugfix, refactor, UI polish, migration, and audit. PM orchestrates specialized agents with feature branches. 98%+ quality gate, runtime verified, visual QA. license: MIT metadata: author: turkeycode version: "9.1.1" website: https://turkeycode.ai allowed-tools: Bash, Read, Write, Edit requires: mcp-servers: - aimem # Required for agent coordination and cross-project learning

Turkey Build

⚡ Quick Start (REQUIRED)

First time in a new project? Two-step process:

Step 1: Bootstrap

/turkey-build
# Creates: .claude/settings.local.json, git init, branches
# Then tells you to restart

Step 2: Restart & Build

# Exit Claude Code (Ctrl+C or /exit)
claude
/turkey-build
# Now agents have permissions and will work!

Why restart? Permissions load when Claude Code starts, not mid-session.

If agents fail with "Permission denied" → You skipped the restart. Exit and re-run claude.

STOP. READ THESE FILES FIRST.

Before writing ANY code, you MUST read these files in order:

1. references/ORCHESTRATION.md      ← HOW to run the build (features, branches, flow)
2. references/PM-AGENT.md           ← PM orchestrates everything
3. references/AIMEM-INTEGRATION.md  ← HOW agents coordinate through memory

DO NOT use the summaries below as your guide. They are overviews only. DO read the full reference files - they contain the actual implementation details.

What This Skill Does

Say "build me X" → Get production-ready, runtime-verified code.

Key Differentiators:

Feature branches - Branches per feature, not per agent
PM orchestration - PM scopes features, assigns agents, delivers release
Greenfield + iteration - Same workflow for new builds and adding features
aimem coordination - Agents signal completion, PM coordinates
98% quality gate - Conductor won't ship below threshold
Runtime verification - Actually starts the app, hits endpoints, verifies
Bugfix branches - Runtime/Conductor failures get targeted fixes

Execution Flow

🦃 BOOTSTRAP (runs first!)
  - Create .claude/settings.local.json with agent permissions
  - Initialize git, set user config
  - Create main + develop branches
  - Initial commit if empty
         │
         ▼
PM Agent reads requirements (or existing code + request)
         │
         ▼
PM outputs scope.yaml with features:
  - feature/core-daemon (backend)
  - feature/api-server (backend, devops)
  - feature/web-dashboard (designer, frontend)
  - feature/documentation (docs)
         │
         ▼
For each feature (in dependency order):
┌─────────────────────────────────────┐
│  Create feature/* branch            │
│  Dispatch assigned agents           │
│  Agents collaborate on same branch  │
│  Merge to develop when complete     │
└─────────────────────────────────────┘
         │
         ▼
Review Wave (parallel on develop):
  QA, Security, CodeReview, Performance
         │
         ▼
Runtime Verification:
  Start server → Hit endpoints → Verify responses
  Failures → BUGFIX-AGENT → Trace → Fix → Re-verify
         │
         ▼
Data Flow Verification:
  Schema sync → Placeholder detection → Data trace
  Verify real data flows DB → API → UI (no "Unknown")
  Failures → BUGFIX-AGENT → Trace → Fix → Re-verify
         │
         ▼
E2E Browser Testing:
  Open browser → Click through UI → Verify user flows
  Verify ACTUAL DATA displayed (not just elements exist)
  Capture screenshots (3 viewports × all pages × states)
  Failures → BUGFIX-AGENT → Trace → Fix → Re-test
         │
         ▼
Visual QA Analysis:
  Demo Agent calls Visual QA Agent
  Analyze screenshots using Claude vision
  Detect CSS bugs, layout issues, responsive problems
  Critical issues → BUGFIX-AGENT → Fix CSS → Re-capture
         │
         ▼
Conductor Score:
  < 98% → BUGFIX-AGENT → Trace → Fix → Re-score
  ≥ 98% → merge develop → main → tag release

Branch Types

| Prefix | Purpose | Example | |--------|---------|---------| | feature/* | New functionality | feature/user-auth | | bugfix/* | Runtime/Conductor failures | bugfix/empty-dashboard | | develop | Integration branch | Always deployable | | main | Production releases | Tagged versions only |

Seven Modes

Greenfield (New Build)

User: "Build a clipboard history manager"
PM: Scopes 5 features → Builds each → Review → Ship v1.0.0

Iteration (Add Features)

User: "Add date filtering and export to the clipboard manager"
PM: Reads existing code → Scopes 2 new features → Builds each → Review → Ship v1.1.0

Bugfix (Fix Issues)

User: "Cards not showing in battle view"
PM: Creates bugfix branch → BUGFIX-AGENT traces → Finds root cause → Fixes → Verifies

Refactor (Restructure Code)

User: "Clean up the auth code" / "Split the god file"
PM: Analyzes structure → Plans refactor → Restructures → Verifies behavior unchanged

UI Polish (Visual Cleanup)

User: "Make the UI look better" / "Fix the messy CSS"
PM: Visual QA scan → Designer reviews → CSS cleanup → Component polish → Responsive fixes

Migration (Upgrade Dependencies)

User: "Upgrade to React 19" / "Move from Express to Hono"
PM: Audit current → Plan migration → Update incrementally → Test each step → Verify

Audit (Analysis Only)

User: "Review security" / "Check performance" / "Analyze code quality"
PM: Runs relevant agents → Produces report → No code changes unless requested

Same workflow. PM routes to different agents based on task type.

Agent Reference Files

PM reads all agent files. Other agents read their own file.

| Agent | File | Role | |-------|------|------| | PM | PM-AGENT.md | Orchestrator - Scopes features, assigns agents | | Discovery | DISCOVERY-AGENT.md | Requirements → scope.yaml (called by PM) | | Designer | DESIGNER-AGENT.md | Design tokens, component specs | | Backend | BACKEND-AGENT.md | API, database, services | | Frontend | FRONTEND-AGENT.md | UI components | | Docs | DOCS-AGENT.md | README, API docs, CLAUDE.md | | DevOps | DEVOPS-AGENT.md | Docker, CI/CD | | QA | QA-AGENT.md | Unit/integration tests | | Security | SECURITY-AGENT.md | Vulnerability scanning | | Code Review | CODE-REVIEW-AGENT.md | Quality analysis | | Performance | PERFORMANCE-AGENT.md | Optimization | | Demo | DEMO-AGENT.md | User perspective critique (calls Visual QA) | | E2E | E2E-AGENT.md | Browser testing - Playwright, real user flows, screenshot capture | | Visual QA | VISUAL-QA-AGENT.md | Screenshot analysis - Claude vision, CSS bugs, layout issues | | Data Flow | DATA-FLOW-VERIFICATION.md | Schema sync, placeholder detection, data trace | | Bugfix | BUGFIX-AGENT.md | Systematic debugging - Reproduce, trace, isolate, fix, verify | | Conductor | CONDUCTOR-AGENT.md | Quality gate (98/100 with visual QA dimension) |

All files in references/ directory.

Agent Tool Requirements

Each agent requires specific tools based on their responsibilities:

| Agent | Required Tools | Purpose | |-------|---------------|---------| | PM | Bash, Read, Write, Edit | Git operations, file management, orchestration | | Discovery | Read | Analyze existing code for iteration mode | | Designer | Write, Edit | Create design tokens, semantic registry | | Backend | Bash, Read, Write, Edit | Run migrations, create APIs, database setup | | Frontend | Read, Write, Edit | Read registry/contracts, build UI components | | Docs | Read, Write, Edit | Create README, API docs, CLAUDE.md | | DevOps | Bash, Write, Edit | Docker setup, CI/CD configuration | | QA | Bash, Read, Write, Edit | Run tests, create test files | | Security | Bash, Read | Run security scans, analyze code | | Code Review | Read, Edit | Analyze and refactor code | | Performance | Bash, Read, Edit | Profile code, optimize queries | | Demo | Read | Analyze from user perspective | | E2E | Bash, Read, Write, Edit | Playwright setup, browser tests, add data-testid | | Bugfix | Bash, Read, Write, Edit | Trace data flow, diagnose root cause, apply fixes | | Conductor | Read | Score quality, make ship decisions |

Critical Rules

PM runs first and last - PM orchestrates the entire build
Feature branches, not agent branches - feature/user-auth not feature/backend
Multiple agents per feature - Designer + Frontend collaborate on same branch
Bugfix branches for failures - Don't rebuild, fix targeted issues
Iteration reads existing code - PM understands what exists before adding
98% threshold - Conductor iterates until quality passes
Runtime verification - "Compiles" ≠ "Works"
Data flow verification - No "Unknown" or placeholder data in UI

File Structure

auto-app-builder/
├── SKILL.md                          # This file (entry point)
├── README.md
├── LICENSE
└── references/
    ├── ORCHESTRATION.md              # ← READ FIRST
    ├── PM-AGENT.md                   # ← PM orchestrates
    ├── AIMEM-INTEGRATION.md          # ← Agent coordination
    ├── DISCOVERY-AGENT.md
    ├── DESIGNER-AGENT.md
    ├── BACKEND-AGENT.md
    ├── FRONTEND-AGENT.md
    ├── QA-AGENT.md                   # ← Includes schema sync checks
    ├── DEVOPS-AGENT.md
    ├── DEMO-AGENT.md                 # ← Calls Visual QA for screenshot analysis
    ├── E2E-AGENT.md                  # ← Browser testing + screenshot capture
    ├── VISUAL-QA-AGENT.md            # ← NEW: Claude vision screenshot analysis
    ├── DATA-FLOW-VERIFICATION.md     # ← Schema sync, placeholder detection
    ├── BUGFIX-AGENT.md               # ← Systematic debugging methodology
    ├── CONDUCTOR-AGENT.md            # ← Quality gate with Visual QA dimension
    ├── SECURITY-AGENT.md
    ├── CODE-REVIEW-AGENT.md
    ├── PERFORMANCE-AGENT.md
    ├── DOCS-AGENT.md
    ├── MEMORY-COORDINATION.md
    ├── QA-SCORING.md
    └── RUNTIME-VERIFICATION.md

License

MIT • TurkeyCode.ai

Constructeur d'applications multi-agents

Recommandé pour