Harness Creator

English | 中文

Generate a customized Claude Code harness framework with Fusion Architecture: GAN-inspired multi-agent system + Domain Specialists.

---

What is Harness Creator?

Harness Creator is a meta-framework that generates customized .claude/ harness structures for AI-assisted software development. It implements Anthropic's best practices for long-running agent tasks with an enhanced Fusion Architecture.

Key Features

• Fusion Architecture: GAN-inspired core + Domain Specialists
• One-Click Customized Harness: Auto-detect project type and domain
• Domain Templates: Pre-defined templates for web, game, data science, etc.
• Dynamic Specialist Generation: Support for any new domain
• Sprint Contract Mechanism: Negotiate "done" criteria before implementation
• Context Reset Protocol: Structured handoffs for long-running tasks

---

Three Layers

Layer	Agents	Responsibility
Decision	Planner, architect-lead	Product & Architecture decisions
Execution	Generator, [domain]-dev	Implement features
Evaluation	Evaluator	Test and grade implementation

Why Separation Matters

Models cannot reliably evaluate their own work. When a single agent both builds and reviews:

• Praises mediocre output as "complete"
• Skips edge cases and error handling
• Misses integration issues
• Accepts incomplete solutions

The solution: Separate generation and evaluation, delegate to specialists.

---

Quick Start

Installation

npx skills add https://github.com/fanlw0816/harness-creator --skill harness-creator

Usage

/harness-creator [target-directory]

Example Session

> /harness-creator ./my-game

## Project Analysis

**Path:** ./my-game
**Type:** Game Development (Unity detected)

## Recommended Configuration

Core Agents:
  Planner        - Product planning
  Generator      - Coordinate execution
  Evaluator      - Quality assessment
  architect-lead - Architecture decisions

Specialists:
  [√] gameplay-dev    Game logic, mechanics, AI
  [√] graphics-dev    Rendering, shaders, effects
  [√] audio-dev       Sound system, music
  [√] ui-dev          Game UI, HUD, menus

Options:
  A) Use recommended
  B) Adjust selection
  C) Add custom specialist

> A

Harness generated at ./my-game/.claude/

---

Domain Templates

Supported Domains

Domain	Specialists
Web Development	frontend-dev, api-dev, database-dev, devops-dev
Game Development	gameplay-dev, graphics-dev, audio-dev, ui-dev
Data Science	data-engineer, ml-engineer, data-analyst
Mobile App	ios-dev, android-dev, backend-dev
Custom	Dynamically generated based on project structure

Template Structure

domain-templates/
├── web-development/
│   ├── template.yaml
│   └── specialists/
│       ├── frontend-dev.yaml
│       └── api-dev.yaml
├── game-development/
│   └── ...
└── _dynamic/                    # Dynamic generation
    ├── specialist-template.yaml
    └── detection-rules.yaml

Adding New Domains

1. Create template directory: domain-templates/your-domain/
2. Define template.yaml with detection rules
3. Add specialist definitions in specialists/*.yaml

---

Sprint Contract

The Contract Lifecycle

1. PROPOSE     Generator proposes contract (with specialist assignments)
      |
      v
2. REVIEW      architect-lead reviews architecture
      |         Evaluator reviews testability
      v
3. NEGOTIATE   Back-and-forth until agreement
      |
      v
4. APPROVE     All parties sign off
      |
      v
5. IMPLEMENT   Generator coordinates specialists
      |
      v
6. EVALUATE    Evaluator grades against contract
      |
      v
7. ITERATE     If FAIL: fix issues, re-evaluate

Contract Structure

# Sprint Contract: [Feature Name]

## Scope
- What, In Scope, Out of Scope

## Architecture Decision
- Specialists Involved
- Cross-Domain Boundaries

## Testable Behaviors
- [ ] B1.1: [Behavior] | Owner: [specialist]

## Acceptance Criteria
| ID | Criterion | Pass | Fail | Priority | Owner |

## Responsibility Matrix
| Criterion | Responsible | Fallback |

---

Generated Harness Structure

<target-dir>/
├── .claude/
│   ├── settings.json
│   ├── agents/
│   │   ├── planner.md
│   │   ├── generator.md          # + Agent tool
│   │   ├── evaluator.md
│   │   ├── architect-lead.md
│   │   └── [domain]-dev.md
│   ├── skills/
│   │   ├── start/
│   │   ├── checkpoint/
│   │   ├── resume/
│   │   ├── sprint-contract/
│   │   └── [domain-specific]/
│   ├── hooks/
│   ├── rules/
│   ├── evaluation/
│   └── docs/
├── production/
│   ├── session-state/
│   └── session-logs/
└── CLAUDE.md

---

Key Principles

1. Generator-Evaluator Separation: Never the same agent
2. Sprint Contract First: Negotiate "done" before building
3. Domain Specialists: Delegated by Generator, specialized implementation
4. Context Reset: Use handoff artifacts for long tasks
5. Concrete Criteria: Convert subjective to gradable

---

Reference Architecture

harness-creator/
├── SKILL.md
├── references/
│   ├── base/
│   │   ├── core/
│   │   │   ├── agents/            # planner, generator, evaluator
│   │   │   ├── skills/            # sprint-contract
│   │   │   └── templates/
│   │   ├── progress/              # start, checkpoint, resume
│   │   ├── evaluator/             # evaluate-feature, evaluate-code
│   │   ├── safety/                # hooks, rules
│   │   └── settings/
│   ├── brainstorm/
│   ├── customizer/
│   └── domain-templates/
│       ├── web-development/
│       ├── game-development/
│       ├── data-science/
│       └── _dynamic/

---

Reference

Design based on Anthropic: Harness Design for Long-Running Apps

---

License

MIT License - See LICENSE for details.