prd-generator

PRD Generator

Purpose

Transform a rough product idea into a comprehensive, AI-ready Product Requirements Document (PDF) through targeted questions and structured output.

Execution Logic

Check $ARGUMENTS first to determine execution mode:

If $ARGUMENTS is empty or not provided:

Respond with: "prd-generator loaded, describe your product idea"

Then wait for the user to provide their product concept in the next message.

If $ARGUMENTS contains content:

Proceed immediately to Task Execution (skip the "loaded" message).

Task Execution

1. MANDATORY: Read Reference Files FIRST

BLOCKING REQUIREMENT — DO NOT SKIP THIS STEP

Before doing ANYTHING else, use the Read tool to read:

• ./references/prd_template.md

This template defines the exact structure your PRD must follow. DO NOT PROCEED to Step 2 until you have read this file.

2. Skip Business Context

This skill intentionally DOES NOT read FOUNDER_CONTEXT.md. PRDs are standalone documents that should contain all necessary context within them.

3. Analyze Initial Input

From the user's initial description, extract what's available:

• Product name or working title
• Core problem being solved
• Target users/audience
• Key features mentioned
• Technical preferences (if any)
• Constraints or requirements (if any)

4. Ask Clarifying Questions

Use AskUserQuestion tool to gather missing information. Ask up to 7 questions maximum, but fewer is better — stop as soon as you have enough to build a comprehensive PRD.

Question Bank (priority order):

Question strategy:

• Ask 2-4 questions per batch using AskUserQuestion
• If the first batch answers provide enough detail, stop asking
• Never ask more than 7 questions total
• Group related questions when possible

5. Generate the PRD

Using the template structure from ./references/prd_template.md, create a complete PRD:

1. Fill every applicable section from the template
2. Be specific — vague requirements produce vague code
3. Write acceptance criteria for every feature — make them testable
4. Prioritize ruthlessly — P0 should be 30-40% of features
5. The "Implementation Notes for AI" section is mandatory — this is what makes it AI-ready

6. Save and Convert to PDF

Step 6a: Create output folder

mkdir -p ./prd_outputs/[Project Name]/

Use the product name with spaces, e.g., ./prd_outputs/Churn Prevention Tool/

Step 6b: Save markdown file Write the PRD content to:

./prd_outputs/[Project Name]/[project_name]_PRD.md

Use snake_case for the filename, e.g., churn_prevention_tool_PRD.md

Step 6c: Convert to PDF Run:

npx md-to-pdf "./prd_outputs/[Project Name]/[project_name]_PRD.md"

This creates [project_name]_PRD.pdf in the same folder.

7. Confirm Output

Tell the user:

• Where the PDF is saved (full path)
• Where the markdown source is saved
• Brief summary of what's in the PRD

Writing Rules

Core Rules

• Every feature MUST have testable acceptance criteria
• Use specific numbers, not vague terms ("loads in <2s" not "loads quickly")
• P0 features should be 30-40% of total features — if everything is P0, nothing is
• Data models must include field types and relationships
• API specs must include request/response examples

PRD-Specific Rules

• Executive summary: 3-5 sentences maximum
• Problem statement: Must include current state, pain points, and business impact
• User personas: Maximum 3 primary personas — more creates confusion
• Tech architecture: Describe data flow in plain English — AI tools interpret this better than complex diagrams
• Implementation Notes for AI section: This is mandatory, never skip it

Format Rules

• Use markdown headers consistently (# for title, ## for sections, ### for subsections)
• Use tables for structured data (metrics, data models, API specs)
• Use code blocks for JSON examples and technical specs
• Use checkboxes for acceptance criteria

Output Format

The PRD follows the structure in ./references/prd_template.md. Here's a condensed example:

References

This file MUST be read using the Read tool before task execution (see Step 1):

Why this matters: The template ensures every PRD follows a consistent, comprehensive structure that AI coding tools can parse and implement. Skipping the template results in incomplete PRDs that miss critical sections.

Quality Checklist (Self-Verification)

Pre-Execution Check

• I read ./references/prd_template.md before starting
• I have the template structure in context

Question Check

• I asked 7 or fewer questions total
• I only asked questions where information was genuinely missing
• Questions were batched (2-4 per AskUserQuestion call)

PRD Content Check

• Executive summary is 3-5 sentences
• Every feature has acceptance criteria (checkboxes)
• P0 features are ~30-40% of total (not everything)
• Data models include field types
• API specs include request/response examples
• "Implementation Notes for AI" section is complete

Output Check

• Markdown file saved to ./prd_outputs/[Project Name]/
• PDF generated via npx md-to-pdf
• User informed of file locations

If ANY check fails → fix before completing.

Defaults & Assumptions

Use these unless the user specifies otherwise:

• Document version: 1.0
• Status: Draft
• Author: PRD Generator
• Tech stack: Modern web (React + Node.js + PostgreSQL) unless specified
• Hosting: Cloud-native (Vercel/Railway/AWS) unless specified
• Auth: Third-party (Clerk/Auth0) unless building custom
• Priority split: ~35% P0, ~40% P1, ~25% P2
• User personas: Maximum 3 unless complexity demands more
• API style: REST unless GraphQL is specified

Document any assumptions made in the PRD output.