Harness Loop (Ralph Wiggum Technique)

Run Claude Code in an external bash loop for autonomous, iterative development.

The Harness Loop implements the Ralph Wiggum technique for fully autonomous development by running Claude Code in an external bash loop. Each iteration, Claude reads specs, works on tasks, logs progress, and continues until all acceptance criteria are met.

Overview

Key benefits:

• •Autonomous execution: Claude works through tasks without manual intervention
• •Progress tracking: Acceptance criteria checkboxes track completion
• •Iteration logs: Every iteration creates a detailed report
• •Graceful completion: Loop stops when "## COMPLETED" marker is added to PLAN.md

Quick Start

# 1. Initialize a feature folder
devorch harness init my-feature

# 2. Generate specs (Claude investigates and writes detailed specs)
devorch harness create-specs my-feature

# 3. Review and edit PLAN.md to add tasks

# 4. Start the autonomous loop
devorch harness loop my-feature

Available Commands

devorch harness init <feature>

Creates the feature folder structure with templates.

devorch harness init auth-system

Creates:

devorch/harness/auth-system/
├── PROMPT.md      # Prompt sent to Claude each iteration
├── PLAN.md        # Tasks with acceptance criteria
├── specs/         # For specification files
└── logs/          # Iteration reports

devorch harness create-specs <feature>

Boots Claude interactively to investigate the codebase and write detailed specifications.

devorch harness create-specs auth-system

# Or run in headless mode (no user interaction)
devorch harness create-specs auth-system --headless

Claude will:

1. •Ask clarifying questions about the feature (interactive mode)
2. •Explore the existing codebase structure
3. •Identify relevant patterns and conventions
4. •Write specification files in specs/
5. •Update PLAN.md with tasks

devorch harness loop <feature>

Runs Claude in an external bash loop until completion.

# Basic usage
devorch harness loop auth-system

# With iteration limit
devorch harness loop auth-system --max-iterations 20

# Show all message types (tool calls, etc.)
devorch harness loop auth-system --verbose

Options:

Loop behavior:

1. •Reads PROMPT.md and sends to Claude
2. •Claude works on the first incomplete task
3. •Claude creates iteration log in logs/
4. •Claude updates PLAN.md with progress
5. •Loop checks for "## COMPLETED" marker
6. •Repeats until complete or max iterations reached

Stopping the loop:

• •Press Ctrl+C to stop gracefully after current iteration
• •Add ## COMPLETED to PLAN.md to signal completion

devorch harness list

Lists all existing features with their status.

devorch harness list

Output:

Found 3 feature(s):

  [DONE] auth-system
     3 specs, 12 iterations

  [....] payment-flow
     2 specs, 5 iterations

  [    ] new-feature
     no specs, no iterations

devorch harness status <feature>

Shows detailed progress for a feature.

devorch harness status auth-system

Workflow

1. Initialize Feature

devorch harness init my-feature

This creates the folder structure and template files.

2. Create Specifications

devorch harness create-specs my-feature

When prompted, describe what the feature should do. Claude will investigate your codebase and create detailed specs in the specs/ folder.

3. Define Tasks in PLAN.md

Edit PLAN.md to define tasks with acceptance criteria:

# Plan: my-feature

## Status: IN_PROGRESS

## Tasks

### Task 1: Create database schema
**Acceptance Criteria:**
- [ ] Add users table with email, password_hash columns
- [ ] Add sessions table with user_id, token, expires_at
- [ ] Create migration file

### Task 2: Implement authentication endpoints
**Acceptance Criteria:**
- [ ] POST /auth/register - create new user
- [ ] POST /auth/login - authenticate and return token
- [ ] POST /auth/logout - invalidate session
- [ ] Add input validation for all endpoints

4. Run the Loop

devorch harness loop my-feature

Claude will:

1. •Read specs and PLAN.md
2. •Find the first task with unchecked criteria
3. •Work on that task only
4. •Create an iteration log
5. •Update PLAN.md checkboxes
6. •Repeat until all tasks complete

5. Monitor Progress

Check progress anytime:

devorch harness status my-feature

Or review iteration logs in devorch/harness/my-feature/logs/.

File Formats

PROMPT.md

The prompt sent to Claude each iteration. Created at init time with context training path baked in.

PLAN.md

Tracks tasks and progress:

# Plan: {{FEATURE_NAME}}

## Status: IN_PROGRESS

## Tasks

### Task 1: [Task Name]
**Acceptance Criteria:**
- [ ] Criterion 1
- [ ] Criterion 2

### Task 2: [Task Name]
**Acceptance Criteria:**
- [ ] Criterion 1
- [ ] Criterion 2

Iteration Logs

Created in logs/iteration-NNN.md:

# Iteration 005

**Timestamp:** 2024-01-15T14:30:00Z
**Task:** Task 2 - Implement authentication endpoints
**Status:** COMPLETE

## Changes Made
- `src/routes/auth.ts` - Created - Added login/register/logout endpoints
- `src/middleware/auth.ts` - Created - Token verification middleware

## Approach
Implemented JWT-based authentication following existing patterns in the codebase.

## Verification
- [x] POST /auth/register - create new user: Works, tested with curl
- [x] POST /auth/login - authenticate and return token: Returns valid JWT
- [x] POST /auth/logout - invalidate session: Removes session from DB

## Next Steps
Move to Task 3: Add middleware to protected routes

Best Practices

Writing Good Acceptance Criteria

• •Be specific: "Add email validation" → "Validate email format using regex, return 400 for invalid"
• •Be testable: Each criterion should be verifiable
• •Be atomic: One thing per criterion
• •Include edge cases: "Handle duplicate email registration with 409 response"

Organizing Specs

Number spec files for clear ordering (00-, 01-, etc.):

specs/
├── 00-overview.md       # High-level feature description
├── 01-data-model.md     # Database schema, types
├── 02-api-endpoints.md  # REST API specifications
├── 03-business-logic.md # Core logic requirements
└── 99-integration.md    # How components connect (last)

When to Use Harness Loop

Good fit:

• •Multi-step feature implementation
• •Refactoring with clear acceptance criteria
• •Bug fixes requiring multiple changes
• •Adding tests to existing code

Not ideal for:

• •Exploratory work without clear specs
• •One-off quick fixes
• •Tasks requiring human judgment at each step

Learn More

• •Full Documentation: docs/user-guide/harness-loop.md
• •Ralph Wiggum Technique: https://ghuntley.com/ralph/