Harness Creator
Generate a customized Claude Code harness framework for your project with Fusion Architecture: GAN-inspired multi-agent system + Domain Specialists.
Quick Start
/harness-creator [target-dir]- With argument: Use specified directory
- Without argument: Ask for target directory
- New project (0→1): Full brainstorm flow
- Existing project (1→N): Auto-analyze and quick setup
Fusion Architecture
Core Pattern: GAN-Inspired Multi-Agent
┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ Planner │ ──► │ Generator │ ──► │ Evaluator │
│ (Plans) │ │ (Implements)│ │ (Grades) │
└─────────────┘ └──────┬──────┘ └─────────────┘
│
spawns specialists
│
┌────────────┼────────────┐
▼ ▼ ▼
┌──────────┐ ┌──────────┐ ┌──────────┐
│ frontend │ │ api │ │ database │
│ -dev │ │ -dev │ │ -dev │
└──────────┘ └──────────┘ └──────────┘Why separate agents? Models cannot reliably evaluate their own work. They praise mediocre output and skip edge cases. Separation enables objective quality assessment.
Three Layers
| Layer | Agents | Responsibility |
|---|---|---|
| Decision | Planner, architect-lead | Product & Architecture decisions |
| Execution | Generator, [domain]-dev | Implement features |
| Evaluation | Evaluator | Test and grade implementation |
Key Components
| Component | Purpose |
|---|---|
| Sprint Contract | Negotiate "done" criteria before implementation |
| Domain Specialists | Specialized implementation agents per domain |
| Evaluation Criteria | Convert subjective judgments to gradable standards |
| Context Reset | Structured handoffs for long-running tasks |
Workflow
1. Parse Arguments
- Validate target path
- Ask user if not provided:
Where should I generate the harness? A) Current directory:. B) Specify path: [input] C) Create new project: [name]
2. Analyze Target Directory
Auto-detect:
- Project type:
package.json→ Node.js,pyproject.toml→ Python - Architecture:
services/→ microservices,apps/+packages/→ monorepo - Tech stack: Read dependencies from config files
- Module boundaries: Detect
src/,apps/,packages/ - Domain type: Match against domain templates
Generate analysis report → User confirms or adjusts
3. Domain Detection (NEW in v2.0)
Analyze project structure
│
▼
┌─────────────────────────────────────┐
│ Match against domain templates: │
│ │
│ • web-development (frontend+api+db)│
│ • game-development (gameplay+gfx+audio)│
│ • data-science (ml+data+analysis)│
│ • mobile-app (ios+android)│
│ • ... │
└─────────────────────────────────────┘
│
├─────────────┬─────────────┐
▼ ▼ ▼
Matchfound No match User specifies
│ │ │
▼ ▼ ▼
Use template Dynamic gen Custom config4. Specialist Selection
检测到项目类型: Game Development (Unity)
推荐的配置:
核心 Agents (固定):
Planner - 产品规划
Generator - 协调执行 (+Agent工具)
Evaluator - 评估实现
architect-lead - 架构决策
Specialists (可调整):
[√] gameplay-dev 游戏逻辑、机制、AI
[√] graphics-dev 渲染、shader、特效
[√] audio-dev 音效、音乐系统
[ ] physics-dev 物理模拟、碰撞检测
[√] ui-dev 游戏UI、HUD、菜单
操作:
A) 直接使用推荐
B) 调整选择
C) 添加自定义 specialist
D) 查看详情5. Generate Harness (Parallel Execution)
Use subagents to generate components in parallel for speed:
User confirms generation│
├─► Subagent 1: Core agents (planner, generator, evaluator, architect-lead)
│
├─► Subagent 2: Domain specialists (from template or dynamic)
│
├─► Subagent 3: Evaluation system (skills, templates, docs)
│
├─► Subagent 4: Progress system (skills, hooks, templates)
│
├─► Subagent 5: Safety system (hooks, rules, docs)
│
├─► Subagent 6: Domain-specific skills and rules
│
└─► Main thread: Integrate results + generate settings.json6. Post-Generation Guidance
✅ Harness generated at [target-dir]/.claude/
Next steps:
1. cd [target-dir]
2. claude
3. Run /startDomain Template System
Template Location
domain-templates/
├── web-development/ # Web应用模板│ ├── template.yaml
│ └── specialists/
│ ├── frontend-dev.yaml
│ ├── api-dev.yaml
│ ├── database-dev.yaml
│ └── devops-dev.yaml
│
├── game-development/ # 游戏开发模板
│ ├── template.yaml
│ └── specialists/
│ ├── gameplay-dev.yaml
│ ├── graphics-dev.yaml
│ ├── audio-dev.yaml
│ └── ui-dev.yaml
│
├── data-science/ # 数据科学模板│ └── ...
│
└── _dynamic/ # 动态生成│ ├── specialist-template.yaml
└── detection-rules.yamlTemplate Structure (template.yaml)
template:
name: game-development
version: 1.0.0
# Detection rules
detection:
files: ["*.unity", "*.gd"]
directories: ["Assets/", "Scripts/"]
keywords: ["game", "unity"]
# Core agents (fixed)
core_agents:
- planner
- generator
- evaluator
- architect-lead
# Domain specialists
specialists:
- name: gameplay-dev
source: ./specialists/gameplay-dev.yaml
- name: graphics-dev
source: ./specialists/graphics-dev.yamlSpecialist Definition
name: gameplay-dev
description: "Specialist for gameplay logic"
delegation:
mode: spawned_by_generator
contract_required: true
domain_scope:
directories:
- "Scripts/Gameplay/"
- "Scripts/AI/"responsibilities:
- Implement game mechanics
- Create state machines
- Implement AI behavior
anti_patterns:
- "Don't put rendering code in gameplay scripts"Generated Harness Structure
<target-dir>/
├── .claude/
│ ├── settings.json
│ ├── agents/
│ │ ├── planner.md
│ │ ├── generator.md # + Agent 工具
│ │ ├── evaluator.md
│ │ ├── architect-lead.md
│ │ └── [domain]-dev.md # 领域 specialist
│ ├── skills/
│ │ ├── start/
│ │ ├── checkpoint/
│ │ ├── resume/
│ │ ├── sprint-contract/
│ │ └── [domain-specific]/ # 领域特定 skills
│ ├── hooks/
│ ├── rules/
│ ├── evaluation/
│ └── docs/
├── production/
│ ├── session-state/
│ └── session-logs/
└── CLAUDE.mdKey Principles
- Generator-Evaluator Separation: Never the same agent
- Sprint Contract First: Negotiate "done" before building
- Domain Specialists: Delegated by Generator, specialized implementation
- Context Reset: Use handoff artifacts for long tasks
- Concrete Criteria: Convert subjective to gradable
Reference
Design based on Anthropic: Harness Design for Long-Running Apps