Agent-Skills.md

Agent Skills: Ralph PRD Creation

Create Product Requirements Document (PRD) and setup for Ralph autonomous loop. Use when user runs /create-prd command, wants to set up a project for Ralph, mentions "ralph setup", "create prd", "product requirements", or needs to generate tasks for autonomous development.

UncategorizedID: joaquimscosta/arkhe-claude-plugins/ralph-prd

Repository

joaquimscosta/arkhe-claude-plugins
joaquimscostaLicense: MIT
71

Install this agent skill to your local

pnpm dlx add-skill https://github.com/joaquimscosta/arkhe-claude-plugins/tree/HEAD/plugins/ralph/skills/ralph-prd

Skill Files

Browse the full folder contents for ralph-prd.

Download Skill

Loading file tree…

plugins/ralph/skills/ralph-prd/SKILL.md

Skill Metadata

Name
ralph-prd
Description
Create Product Requirements Document (PRD) and setup for Ralph autonomous loop. Use when user runs /create-prd command, wants to set up a project for Ralph, mentions "ralph setup", "create prd", "product requirements", or needs to generate tasks for autonomous development.

Ralph PRD Creation

Interactive wizard to create Product Requirements Document and Ralph project setup with task sets.

Quick Start

/create-prd                    # Start interactive wizard
/create-prd "Build a todo app" # Start with description

What It Creates

project/
├── PROMPT.md                          # Instructions for each iteration
├── ralph.sh                           # Loop runner (executable)
└── .ralph/
    ├── current-taskset -> tasksets/initial  # Symlink to active taskset
    └── tasksets/
        └── initial/                   # Your task set (named during setup)
            ├── tasks.json             # Task list (JSON)
            ├── prd.md                 # Requirements (Markdown)
            ├── memories.md            # Persistent learnings
            ├── config.json            # Ralph settings
            └── activity.log           # Iteration log (empty)

Discovery Questions

Ask these discovery questions:

  1. Task Set Name - What should this collection be called? (default: "initial")
  2. Problem - What problem are you solving?
  3. Audience - Who is the target user?
  4. Features - What are the 3-5 core features?
  5. Tech Stack - What technologies to use?
  6. Architecture - Monolith, microservices, etc.?
  7. UI/UX - Visual requirements and preferences?
  8. Auth - Authentication needs?
  9. Integrations - Third-party services?
  10. Success Criteria - How do we know it's done?

Question Specificity for Destructive Operations

When discovery questions involve deleting files, removing dependencies, or other destructive changes:

  • Always include full paths of items being deleted or modified
  • State size/scope — file count, line count, component count
  • Explicitly state what will NOT be affected
  • Frame as confirmation, not open question — "Deleting X, Y, Z. Proceed?" not "What should be deleted?"

See EXAMPLES.md for vague-vs-specific anti-patterns.

Task Generation

Convert features into atomic tasks with categories: setup, feature, integration, styling, testing, verification. See WORKFLOW.md for task format and categories.

Final verification task: Always include a final task with verificationTier: "visual" (UI projects) or "api" (API-only projects) that verifies the complete application works end-to-end. This catches issues that individual task verification misses.

Workflow

See WORKFLOW.md for detailed discovery flow.

Examples

See EXAMPLES.md for PRD examples.

Troubleshooting

See TROUBLESHOOTING.md for common issues.

Templates (MANDATORY)

You MUST read and use these templates when generating output files. Do NOT write simplified versions from memory — the loop depends on exact field names and signal formats.

  • prompt.md.template - CRITICAL: Contains the RALPH_COMPLETE: signal in Step 5. Read this template with the Read tool and fill in {{placeholders}}. Never hand-write PROMPT.md.
  • tasks.json.template - CRITICAL: Tasks use "passes": false / "passes": true (NOT "status": "pending" / "status": "done"). The completion signal checks passes.
  • prd.md.template - PRD document structure
  • config.json.template - Ralph configuration
  • memories.md.template - Learnings file

Why This Matters

The ralph.sh loop script checks for RALPH_COMPLETE: in the output to stop. The prompt.md.template Step 5 tells Ralph to emit this signal when all tasks have passes: true. If PROMPT.md is hand-written without Step 5, or tasks.json uses a different field name, the loop runs forever.

ralph.sh

Always copy ralph.sh from this plugin's scripts/ralph.sh directory. It contains required flags (--dangerously-skip-permissions, --disallowedTools) that are essential for non-interactive execution.

After Setup

./ralph.sh 20    # Start autonomous loop

Creating Additional Task Sets

After initial setup, create more task sets:

/ralph taskset new "auth-feature"    # Create new task set
/ralph taskset list                  # See all task sets
/ralph taskset switch "auth-feature" # Switch to it
Ralph PRD Creation Skill | Agent Skills