Ralph PRD Skill for Claude Code

A Claude Code skill for generating Product Requirements Documents and running the Ralph autonomous implementation loop.

What is Ralph?

Ralph is an autonomous coding workflow that:

1. Reads your PRD (Product Requirements Document)
2. Implements one task at a time
3. Runs tests to verify the implementation
4. Commits working code automatically
5. Repeats until all tasks are complete

What's Included

Component	Description
SKILL.md	The PRD generator skill for Claude Code
ralph.sh	The autonomous implementation loop script

Prerequisites

• Claude Code installed and configured
• A project directory where you want to use Ralph

Installation

Run this command in your project directory:

curl -fsSL https://raw.githubusercontent.com/mykcryptodev/ralph-prd/master/install.sh | bash

This installs:

• .claude/skills/prd/SKILL.md - the PRD generator skill
• ralph.sh - the autonomous loop script

Updating

Run the same install command to update to the latest version.

Uninstallation

To remove Ralph from your project:

rm -rf .claude/skills/prd ralph.sh

Any PRD.md or progress.txt files are not affected.

Usage

Creating a PRD

Use the /prd command or just ask Claude to "create a prd":

/prd add user authentication

Or simply describe what you want to build:

create a prd for a REST API for managing tasks

The skill will:

1. Ask clarifying questions if needed
2. Generate a structured PRD.md in your project root
3. Create a progress.txt file for tracking learnings

Running the Autonomous Loop

After the PRD is created, start Ralph:

./ralph.sh

Ralph will iterate through each user story in your PRD, implementing them one at a time until all tasks are marked complete.

You can optionally specify a maximum number of iterations:

./ralph.sh 25  # Run up to 25 iterations (default is 50)

What Gets Created

When you use the /prd command, these files are created in your project:

File	Description
PRD.md	The Product Requirements Document with user stories and acceptance criteria
progress.txt	Tracks learnings and iteration history for Ralph
ralph.sh	Executable script that runs the autonomous implementation loop

How Ralph Works

Each iteration of Ralph:

1. Reads PRD.md - Finds the first incomplete task (marked [ ])
2. Checks progress.txt - Reviews learnings from previous iterations
3. Implements the task - Makes the necessary code changes
4. Runs tests - Verifies the implementation works
5. On success:
   • Marks the task complete ([x]) in PRD.md
   • Commits the changes
   • Logs what worked in progress.txt
6. On failure:
   • Does NOT mark the task complete
   • Does NOT commit broken code
   • Logs what went wrong in progress.txt

This continues until all tasks are complete or the max iteration limit is reached.

PRD Format

PRDs generated by this skill follow this structure:

# PRD: [Feature Name]

## Introduction
[What this feature does and why it matters]

## Goals
- [Primary and secondary goals]

## User Stories

### US-001: [Story Title]
**Description:** As a [user], I want [capability] so that [benefit].

**Acceptance Criteria:**
- [ ] [Testable criterion]
- [ ] [Another criterion]
- [ ] Typecheck passes (if applicable)

## Non-Goals
- [Things explicitly out of scope]

## Technical Considerations
- [Architecture notes, dependencies, constraints]

Tips for Best Results

1. Keep user stories small - Each should be completable in one iteration
2. Order stories by dependency - Put foundational work first
3. Make acceptance criteria testable - Ralph uses these to verify success
4. Review the PRD before running Ralph - Adjust anything that seems off

Troubleshooting

ralph.sh: Permission denied

Make the script executable:

chmod +x ralph.sh

Ralph keeps failing on the same task

Check progress.txt - it contains learnings from failed attempts. You may need to:

• Simplify the acceptance criteria
• Break the task into smaller steps
• Add technical context to the PRD

Claude Code not finding the /prd command

If /prd is not available:

• Verify the skill is installed: check .claude/skills/prd/SKILL.md exists in your project
• Restart Claude Code after installation
• Reinstall if needed: run the install command again

License

MIT