prd（珠三角）

PRD Creation Skill

Create self-verifying PRDs for autonomous execution. This skill guides you through interviewing users and generating structured PRDs.

Core Philosophy

You are intelligent. These guidelines inform your thinking - they don't constrain it.

• Adapt to the user - Every project is different. Adjust your approach.
• Think independently - You decide what questions to ask and when to stop.
• Verify, don't assume - Challenge assumptions and uncover edge cases.
• Principles over prescriptions - Apply mental models, not rigid scripts.

Important Guardrails

• Do NOT start implementing - Your job is to create the PRD, not write code.
• Ask questions one set at a time - Don't overwhelm with multiple question blocks.
• Always ask about quality gates - This is REQUIRED for every PRD.
• Get explicit approval before generating - Present your understanding first.

For comprehensive reference material, read AGENTS.md. For specific guidance, reference the interview/ and categories/ directories.

The Interview Process

Your goal: Extract enough information to create a PRD that an AI agent can execute successfully.

Phase 1: Identify Task Type

Start by understanding what kind of work this is. Use AskUserQuestion with these categories:

For detailed category guidance, see categories/_overview.md and individual category files.

Phase 2: Brain Dump

Ask the user to share everything they know. Let information flow without imposing structure.

What you need to understand:

• The problem or goal
• Context and background
• Constraints and requirements
• Who benefits and how
• Files, systems, or areas involved
• Sources to reference (docs, APIs, existing code)

For guidance on gathering initial information, see interview/brain-dump.md.

Phase 3: Clarifying Questions

This is where you think independently. Based on what they told you:

• What gaps remain in your understanding?
• What assumptions should you challenge?
• What could go wrong that they haven't mentioned?
• What decisions need their input vs. your judgment?

The number of rounds depends on complexity. Simple tasks: 2-3 rounds. Complex features: 6-10 rounds. You decide when you have enough.

Ask questions one set at a time. After each answer, decide whether to:

• Ask follow-up questions (if answers reveal complexity)
• Ask about a new aspect (if current area is clear)
• Move to confirmation phase (if you have enough context)

For guidance on formulating questions, see interview/clarifying-questions.md.

Lettered Multiple Choice Format

When appropriate, use lettered options to speed up responses:

1. What is the primary goal of this feature?
   A. Improve user onboarding experience
   B. Increase user retention
   C. Reduce support burden
   D. Other: [please specify]

2. Who is the target user?
   A. New users only
   B. Existing users only
   C. All users
   D. Admin users only

This lets users respond with "1A, 2C" for quick iteration.

Quality Gates Question (REQUIRED)

Always ask about quality gates - these follow a two-tiered testing approach:

Tier 1: Story-Level Testing (every story must do this)

• Write ALL tests for the feature being implemented (no tests left out)
• Run lint, typecheck, build verification
• Run ALL tests related to that story: unit tests, integration tests for the code touched, E2E tests if the story has UI
• All tests must be written AND pass before the story is complete

Tier 2: Dedicated Testing Sessions (end of phase or PRD)

• Run the complete test suite to ensure nothing broke
• Verify all E2E tests pass
• For features with UI: comprehensive browser verification

1. What quality commands should run for story-level testing?
   A. pnpm typecheck && pnpm lint && pnpm test --related
   B. npm run typecheck && npm run lint && npm test -- --findRelatedTests
   C. bun run typecheck && bun run lint && bun test --related
   D. Other: [specify your commands]

2. Does this project have E2E tests?
   A. Yes - Playwright
   B. Yes - Cypress
   C. Yes - Other framework
   D. No E2E tests yet (we should add them)

3. For UI stories, should we include browser verification?
   A. Yes, use agent-browser skill to verify visually
   B. No, automated tests are sufficient

Phase 4: Confirm Understanding

Before generating, present your understanding:

• What you think they want (problem + solution)
• Your proposed approach
• Rough phase breakdown
• Testing and verification strategy

Get explicit approval. If they want changes, adapt and re-present.

For guidance on confirmation, see interview/confirmation.md.

Category-Specific Guidance

Each task type has different priorities and workflows. These are thinking frameworks, not templates.

Agent Browser CLI

Use Agent Browser CLI throughout your work for any visual or interactive components.

When to use it:

• Research: Explore documentation, validate approaches
• Features: Verify UI as you build, final validation against spec
• Bugs: Reproduce issues, verify fixes
• QA: Test user flows, visual validation

Quick reference:

agent-browser open http://localhost:3000    # Start session
agent-browser snapshot -i                   # Get interactive elements
agent-browser click @e5                     # Click element
agent-browser fill "[name='email']" "test"  # Fill input
agent-browser screenshot verify.png         # Capture state
agent-browser close                         # Clean up

This is emphasized throughout all category guidance - browser verification is essential for UI work.

Story Structure

Every story in the PRD needs these elements:

Required Components

1. Description - WHAT is this and WHY does it matter? Not HOW.
2. Tasks - Step-by-step instructions. Start with context gathering, end with verification.
3. Acceptance Criteria - How do we know it's done? Specific, verifiable statements.
4. Notes - File paths, patterns to follow, warnings about pitfalls.

Common Story Types

Ralph Loop Optimization

When creating stories, keep in mind that Ralph loops operate with these constraints:

Each iteration starts with no memory - The agent must read .ralph-tui/progress.md to understand prior work. This means:

• Story notes should include file paths, function names, and specific locations
• Reference existing patterns that the agent should follow
• Include gotchas or edge cases discovered during planning

Progress entries must be verbose - The template instructs agents to write detailed progress entries with:

• What was implemented (specific function names, class names)
• Files changed (with descriptions)
• Learnings (patterns, gotchas, architecture insights)
• Quality gate results

Include this context in story notes when relevant:

• Which files are likely to be modified
• What existing patterns to follow
• What the agent should document for future iterations

Verification & Testing

Use the two-tiered testing approach for all PRDs:

Tier 1: Story-Level Testing (REQUIRED for every story)

Every story must:

1. Write tests first - All tests related to the feature being implemented
2. Run story-specific tests - No test is left out from the implementation:

- Lint and typecheck - Unit tests for new code - Integration tests for touched code paths - E2E tests if the story has UI components

1. All tests must pass - Story is not complete until tests are written AND passing

Tier 2: Dedicated Testing Sessions (End of phase/PRD)

Include dedicated testing stories at:

• End of each implementation phase
• Final validation before PRD completion

These sessions:

• Run the complete test suite (not just related tests)
• Ensure all E2E tests pass (full coverage, not just new ones)
• Fix any regressions discovered
• For features: verify E2E tests exist and pass for all user flows

UI work always gets browser verification - if there's a visual component, verify it with Agent Browser CLI.

Output Format

Generate two files in docs/prds/[name]/:

PRD.md (Human-readable)

# [Project Name]

## Overview
[What and why]

## Goals
[Specific outcomes]

## Quality Gates

### Story-Level Testing (every story)
- `[lint command]` - Lint check
- `[typecheck command]` - Type verification
- `[test command --related]` - Run tests related to changed files
- `[build command]` - Build verification

For stories with UI:
- Run E2E tests for the specific feature
- Verify in browser using agent-browser skill

### Dedicated Testing Sessions (end of phase)
- `[full test command]` - Complete test suite
- `[e2e test command]` - All E2E tests
- Fix any regressions before proceeding

## Non-Goals
[Out of scope]

## Technical Approach
[High-level strategy]

## Phases
[Phase breakdown with objectives]

## Testing Strategy
[How verification happens]

## Risks & Mitigations
[What could go wrong and how to handle it]

## Success Criteria
[How we know it's complete]

IMPORTANT: Wrap the final PRD.md content in [PRD]...[/PRD] markers for parsing:

[PRD]
# PRD: [Project Name]

## Overview
...

## Quality Gates
...

## User Stories
...
[/PRD]

prd.json (Machine-readable)

{
  "name": "kebab-case-name",
  "description": "Context for all tasks. Motivation, goals, reference CLAUDE.md.",
  "branchName": "type/feature-name",
  "userStories": [
    {
      "id": "US-001",
      "title": "Short descriptive title",
      "description": "WHAT and WHY - not HOW. Include tasks embedded here:\n\n**Tasks:**\n1. First task\n2. Second task\n3. Third task",
      "acceptanceCriteria": ["First criterion", "Second criterion", "Third criterion"],
      "dependsOn": [],
      "notes": "File paths, patterns, warnings",
      "passes": false
    }
  ]
}

Key points:

• tasks: Embed formatted tasks in description (not available as separate field to template)
• acceptanceCriteria: Use array - Ralph TUI's template engine converts it to string automatically
• Template receives {{acceptanceCriteria}} as a pre-formatted string with checkboxes

After Generation

Provide the user with:

1. What Was Created

• File locations for PRD.md and prd.json
• Number of stories and phases

2. Pre-Flight Check

IMPORTANT: Before starting a Ralph loop, run the pre-flight check:

/ralph-preflight

This verifies:

• No global CLAUDE.md conflict
• Config and template paths are correct
• prd.json structure is valid
• Template variables will map correctly
• Template uses optimized v2 format (no {{recentProgress}}, has gibberish cleanup)
• Template has verbose progress entry format for cross-iteration context

3. Execution Setup

Once prd.json exists and pre-flight passes:

Option A: Simple Branch (recommended for single feature)

# 1. Create feature branch
git checkout -b [branch-name-from-prd]

# 2. Start Ralph in tmux
tmux new-session -d -s ralph-[name] "ralph-tui run --prd docs/prds/[name]/prd.json"
tmux attach-session -t ralph-[name]

# 3. Press 's' to start, then Ctrl+B D to detach

Option B: Git Worktree (for parallel development)

# 1. Create worktree with new branch
git worktree add ../[repo]-[name] -b [branch-name-from-prd]
cd ../[repo]-[name]

# 2. Copy .ralph-tui config if not using shared config
# (worktrees share git but have separate working directories)

# 3. Start Ralph in tmux
tmux new-session -d -s ralph-[name] "ralph-tui run --prd docs/prds/[name]/prd.json"
tmux attach-session -t ralph-[name]

Ask user:

AskUserQuestion: "How do you want to run this?"
├── "Simple branch" (Recommended) - Single feature in current directory
├── "Git worktree" - Parallel development in isolated directory
└── "Just show me the commands" - Manual setup

4. Monitoring Notes

To check progress:

# Reattach to tmux session
tmux attach-session -t prd-[name]

# Detach again (leave it running)
# Press Ctrl+B, then D

# Check progress file
cat .ralph-tui/progress.md

# Check iteration logs
ls -la .ralph-tui/iterations/

Recommended check intervals:

• Simple projects: Every 30-60 minutes
• Complex projects: Every 15-30 minutes initially, then as needed

Understanding BLOCKED states:

• BLOCKED means the AI needs human input
• Reattach and provide the needed information
• The loop will continue once you respond

Where progress is tracked:

5. Troubleshooting

If ralphdui is not found:

# Install ralph-tui globally
cargo install ralph-tui

# Or use via npx (if available)
npx ralph-tui --prd docs/prds/[name]/prd.json

If the loop gets stuck:

1. Reattach to tmux session
2. Check what state it's in
3. Provide input if BLOCKED
4. If truly stuck, Ctrl+C to stop and restart

To start over:

# Reset progress (keeps work, restarts loop)
rm -rf .ralph-tui/
ralphdui --prd docs/prds/[name]/prd.json

6. Important Reminders

• Keep machine running during execution - the loop operates autonomously
• Check back at recommended intervals - BLOCKED tasks need human input
• Human input may be needed for ambiguous situations or decisions
• Work is committed automatically after each completed task
• All tests must pass for a task to be marked COMPLETE
• The loop continues until all tasks are done or it encounters a BLOCKED state

Completion Signals

When executing PRD tasks, use these signals:

Key Reminders

• You are intelligent - use guidelines to think, not to follow blindly
• Adapt to the user and their specific situation
• Ask good questions, but know when to stop
• Agent Browser CLI is essential for visual/interactive work
• The goal is a PRD that an AI agent can execute successfully
• When in doubt, verify. When uncertain, ask.

For comprehensive guidance, read AGENTS.md.