Stats
Actions
Available In
Tags
harness-builder
Auto-configure a production-grade Claude Code harness for any project in under 4 minutes.
Analyzes your codebase, runs an adaptive interview, and generates optimized skills, agents, hooks, and security rules — tailored to your actual tech stack and workflow, not generic templates.
Why harness-builder?
- Hours → 4 minutes. Setting up
.claude/from scratch takes most developers half a day. harness-builder does it in one command. - Evidence-based, not generic. It reads your actual code — imports, config files, CI pipelines — before generating anything. Output reflects your project, not a template.
- Non-destructive by design. Every run backs up existing files. Section markers and a lock file allow surgical updates without touching your customizations.
- Cost-aware from the start. Uses opus only where judgment matters (design, architecture); sonnet for execution. Configurable via
--cost-tier. - Beginner-friendly interview. The 6-round adaptive interview (15–24 questions) adjusts depth to your experience level. Novices get guardrails; power users get full control.
Table of Contents
- Quick Start
- How It Works
- Deep Interview System
- Cost Tiers
- Supported Stacks
- Generated Files
- Commands
- Coexistence and Safety
- Example Output
- Design Principles
- Troubleshooting
- FAQ
- Contributing
- License
Quick Start
# 1. Add from marketplace
claude plugin marketplace add woohyunkim/harness-builder
# 2. Install
claude plugin install harness-builder
# 3. Run setup
/harness-setup
Verify installation:
/harness-status
# Expected: "harness-builder v0.1.0 — active"
How It Works
/harness-setup
│
├── Phase 1: Codebase Analysis
│ Detects: language, framework, database, testing,
│ deployment, CI/CD, package manager, existing config
│ Output: .claude/harness-builder/analysis.json
│
├── Phase 2: User Interview
│ Asks about: project phase, cost tier, automation level,
│ git workflow, testing philosophy, security needs
│ Output: .claude/harness-builder/interview.json
│
├── Phase 3: Configuration Generation
│ Generates: settings.json, hooks, skills, agents,
│ CLAUDE.md, scripts, security rules
│ Output: .claude/* (full harness)
│
└── Phase 4: Verification and Guide
Validates: JSON, permissions, frontmatter, secrets
Generates: .claude/docs/HARNESS-GUIDE.md
Manual vs Automated — Time Comparison
| Task | Manual | harness-builder |
|---|---|---|
| Write CLAUDE.md from scratch | 45–90 min | included in setup |
| Configure hooks.json | 30–60 min | included in setup |
| Define agent profiles | 60–120 min | included in setup |
| Write domain skills | 60–90 min | included in setup |
| Set security deny-lists | 20–30 min | included in setup |
| Re-sync after stack change | 30–60 min | /harness-update (~1 min) |
| Total | 4–7 hours | ~4 minutes |
Deep Interview System
harness-builder does not ask generic questions. The interview runs in 6 rounds with 15–24 questions total, adapting based on your answers.
| Round | Topic | Example Questions |
|---|---|---|
| 1 | Project context | Phase (PoC / growth / stable), team size, deadline pressure |
| 2 | Cost and model policy | Budget per hour, when to use opus vs sonnet |
| 3 | Automation preferences | What to auto-trigger vs require explicit approval |
| 4 | Git and review workflow | Branch strategy, PR review requirements, commit style |
| 5 | Testing philosophy | Coverage targets, TDD preference, E2E tooling |
| 6 | Security posture | Data sensitivity, secret management, audit requirements |
Adaptive depth: If you answer "I'm new to Claude Code" in round 1, rounds 2–6 default to safe, recommended values and skip advanced options. If you indicate experience, every option is surfaced.
Each answer directly maps to one or more generated configuration values. Nothing is inferred from templates — all output is derived from your interview responses and codebase analysis.
