personal-harness

personal-harness

Harness Engineering toolkit for Claude Code — 7 tools based on AI-DLC patterns.

"The bottleneck has moved from writing code to reviewing code, and now to architecture and harness design." — Harness Engineering & AI-DLC

Installation

# Register marketplace (one-time)
/plugin marketplace add Geun-Oh/personal-harness

# Install
/plugin install personal-harness@personal-harness

For local development:

claude --plugin-dir ./personal-harness

Quick Start: From Idea to Production

Step 1 — Define Requirements

Turn your idea into a structured features.json. This replaces vague specs with a machine-verifiable checklist.

# Create project
mkdir my-project && cd my-project && git init

# Create features.json (write it yourself or ask the agent to draft it)
cat > features.json << 'EOF'
{
  "features": [
    {
      "id": "auth-login",
      "name": "Email/Password Login",
      "description": "Login with email and password to receive a JWT",
      "steps": [
        "POST /api/auth/login accepts email and password",
        "Returns JWT token for valid credentials",
        "Sets session token in response cookie",
        "Returns 401 for invalid credentials"
      ],
      "tests": ["tests/auth/login.test.ts"],
      "dependencies": [],
      "passes": false
    }
  ]
}
EOF

# Validate feature definitions against 4 quality criteria
/personal-harness:features-validate

Each feature must be specific (concrete steps), independent (no circular deps), verifiable (pass/fail conditions), and testable (test file paths). The agent can only modify the passes field — everything else is human-defined.

Step 2 — Set Up the Harness

Build the environment so agents can work reliably.

# Create CLAUDE.md (~100 lines, acts as an index)
# Include: build/test commands, key rules, pointers to deeper docs

# Verify document hierarchy (L0-L3)
/personal-harness:pyramid-lint

# Diagnose current maturity level and get action items
/personal-harness:harness-assess

Minimum goal for L1 (Basic Harness):

• CLAUDE.md exists with build/test commands
• Linter configured (eslint, ruff, golangci-lint, etc.)
• Test framework set up
• .gitignore properly configured

Step 3 — Build Features

Implement features from features.json one by one. Hooks run automatically — no manual intervention needed.

# Ask the agent to implement a feature
"Implement the auth-login feature from features.json"

# What happens automatically during coding:
# L1 Hook → syntax/secrets check on every file edit (blocks on CRITICAL)
# L2 Hook → linter/structure check at turn end (blocks on lint failure)
# L3 Hook → test execution at turn end (blocks on test failure)
# Budget Hook → context window obesity warning at turn end

# When done, the agent sets passes: true in features.json

Step 4 — Verify Before PR

# When you run `gh pr create`, L4 automatically suggests a review
# The gate-reviewer agent checks security and architecture

# Confirm all features pass
/personal-harness:features-validate

Step 5 — Maintain Over Time

Agents generate entropy 5-10x faster than humans. Run these periodically.

# Scan for doc-code drift, unused exports, architecture violations
/personal-harness:gardener

# Check if repeated feedback should be promoted to stronger enforcement
/personal-harness:feedback-ladder

# Verify context budget isn't bloated
/personal-harness:context-budget

# Re-assess maturity — plan next level
/personal-harness:harness-assess

Tool Reference

Skills (manual invocation)

Skill	Command	Purpose
Features Validator	/personal-harness:features-validate	Validate feature definitions (specificity, independence, verifiability, testability)
Knowledge Pyramid Linter	/personal-harness:pyramid-lint	Check CLAUDE.md and doc hierarchy (L0-L3) structure
Context Budget Monitor	/personal-harness:context-budget	Analyze token budget allocation (CLAUDE.md ≤5%, rules ≤3%, MEMORY ≤2%)
Harness Maturity Assessor	/personal-harness:harness-assess	Diagnose AI-DLC maturity level (L0-L4) with gap analysis
Feedback Encoding Ladder	/personal-harness:feedback-ladder	Track repeated feedback and suggest stronger encoding levels
Gardener Agent	/personal-harness:gardener	Scan for entropy (doc drift, unused code, architecture violations)

Hooks (automatic execution)