Building Product Requirements

Help users create professional, implementation-ready product requirements. The output is a PRD (the "what" and "why") plus a set of focused RFCs (the "how"), designed so that a coding agent can pick up any single RFC and implement it with minimal hallucination.

Why this structure matters

A PRD without RFCs is too vague for implementation. RFCs without a PRD lack product context and user grounding. The combination gives coding agents:

• Focused context — each RFC addresses one implementation area
• Clear boundaries — explicit scope prevents agents from drifting
• Traceability — every technical decision links back to a user need
• Reduced errors — smaller, well-defined documents produce better code

Workflow overview

Copy this checklist and track progress:

Requirements Progress:
- [ ] Phase 1: Discovery — understand the product vision
- [ ] Phase 2: PRD Draft — write the product requirements document
- [ ] Phase 3: RFC Breakdown — identify implementation areas
- [ ] Phase 4: RFC Drafting — write each RFC
- [ ] Phase 5: Review — validate completeness and consistency
- [ ] Phase 6: Generate Prompt — create implementation prompts for coding agents

Output location and naming

Save all documents in a Product requirements/ directory at the project root. Ask the user if they prefer a different location. If a directory already exists with PRDs or RFCs, use that location instead.

Naming convention:

• Use zero-padded three-digit numbers: 001, 002, 012
• Slugs are lowercase, hyphen-separated, 2-4 words max
• When adding RFCs to an existing project, continue the numbering sequence

Phase 1: Discovery

Before asking questions, read the existing codebase if one exists. Scan the project directory for:

• Architecture patterns (MVVM, Clean Architecture, MVC, etc.)
• Tech stack and frameworks in use
• Existing data models, schemas, and entities
• Services, APIs, and integrations already built
• Naming conventions, DI approach, project structure
• Existing PRDs or RFCs (check for a Product requirements/, docs/, specs/,

or rfcs/ directory)

This context prevents you from producing specs that conflict with the codebase. If a PRD already exists and the user wants to add a feature, skip to Phase 3 (see "Adding an RFC to an existing project" below).

Now interview the user. Adapt your questions to what the codebase and user have already told you — don't re-ask what's obvious from context.

Core questions (ask what's missing, skip what's known):

1. What problem are you solving? Who has this problem, and how painful is it?
2. Who are your users? What are their goals, frustrations, and current workarounds?
3. What does success look like? Measurable criteria — time saved, error rates, adoption targets.
4. What's the MVP scope? What's in for v1, what's explicitly out?
5. Technical constraints? Platform (web/mobile/both), tech stack preferences, third-party services, offline needs.
6. What exists already? Any prior art, mockups, codebases, or competitor references?

Capture answers in your working memory. You'll weave them into the PRD.

#### Adding an RFC to an existing project

If the project already has a PRD and the user wants to add a feature (not start from scratch):

1. Read the existing PRD and all existing RFCs to understand current scope,

data models, glossary, and conventions

1. Skip Phase 2 — the PRD exists
2. Go to Phase 3 to define the new RFC's scope, checking for overlap with

existing RFCs

1. Continue with Phase 4-6 as normal
2. After drafting, update the PRD's Features Overview table with the new RFC

Phase 2: PRD Draft

Use the PRD template. See templates/prd-template.md for the full template with section guidance.

Key sections (every PRD needs these):

1. Overview — project description, goals/objectives, success criteria
2. Problem Statement — the pain, the business context, user pain points
3. User Requirements — personas, job stories (not user stories), functional requirements broken into MVP vs. future
4. Technical Requirements — architecture overview, tech stack, API specs, data models, security, performance targets
5. Design Requirements — interaction patterns, accessibility, platform-specific considerations
6. Success Metrics — quantitative measures tied to goals
7. Timeline — milestones and phases
8. Features Overview — summary table linking each feature to its RFC
9. Glossary — define every domain term; this prevents agent confusion

Writing principles:

• Use job stories ("When [situation], I want to [motivation], so I can [outcome]") rather than user stories. They focus on context and motivation rather than role.
• Be specific about what's in scope and what's not. Ambiguity is where coding agents hallucinate.
• Include edge cases in functional requirements. If you can think of it, write it down.
• Define data models early — field names, types, relationships, constraints. These are the foundation everything else builds on.

Phase 3: RFC Breakdown

Decompose the product into discrete, implementable RFCs. Each RFC should:

• Be implementable independently (or with explicit dependencies noted)
• Have clear inputs and outputs
• Map to a logical area of the codebase
• Be small enough that a coding agent can hold the full context

Common decomposition pattern:

Adapt this to the project. A simple web app might need 3 RFCs. A cross-platform mobile app might need 6+. Always start with RFC-000 as the architecture overview that all other RFCs reference.

Implementation order: After defining all RFCs, determine the order they should be implemented. List them with a one-line justification:

Implementation Order:
1. RFC-000: Architecture — foundation everything else depends on
2. RFC-001: Database & Models — data layer must exist before features
3. RFC-002: Core Feature A — primary user value, validates the architecture
4. RFC-004: API Integration — needed by Feature B
5. RFC-003: Feature B — depends on API integration
6. RFC-005: UI Polish — requires all features to be functional

Include this in the PRD's "Features Overview" section so the implementation agent knows what to build first.

Phase 4: RFC Drafting

Use the RFC template. See templates/rfc-template.md for the full template with section guidance.

Each RFC must include:

1. Overview — what this RFC covers, one paragraph
2. Goals — numbered list, specific and testable
3. Technical Design — architecture, data flow, code structure
4. Implementation Details — platform-specific code samples, file paths, configuration
5. Edge Cases — every failure mode you can anticipate, with handling strategy
6. Testing Strategy — unit tests, integration tests, manual testing checklist
7. Security Considerations — auth, data privacy, permissions
8. Dependencies — libraries, APIs, other RFCs this depends on
9. Related Documents — links back to PRD and sibling RFCs

Writing principles for RFCs:

• Include code samples in the language/framework of the project. Not pseudocode — actual compilable/runnable snippets.
• Specify file paths where code should live (e.g., src/database/schema.ts).
• List explicit exclusions — "Do NOT implement X (that's RFC-003)". This prevents coding agents from scope-creeping.
• Use consistent naming across all RFCs. If the PRD calls it "Idea", every RFC calls it "Idea" (not "Note" or "Entry").

Phase 5: Review

Before finalizing, actively verify each item — don't just eyeball it.

1. Traceability check: Read every JS-XXX job story in the PRD. For each one,

search the RFC files to confirm at least one RFC references it. List any orphaned stories.

1. Data model consistency: Compare entity field names, types, and relationships

across the PRD and every RFC that references them. Flag any mismatches (e.g., PRD says categoryId: UUID but RFC-003 says category: String).

1. Scope overlap check: For each RFC, read its "Scope Boundaries" section.

Flag any capability that appears in more than one RFC's "This RFC covers" list.

1. Exclusion check: Verify each RFC has an explicit "This RFC does NOT cover"

section with cross-references to sibling RFCs.

1. Terminology consistency: Pick the glossary terms from the PRD and search

all RFC files for variant spellings or synonyms (e.g., "Note" vs "Idea", "Category" vs "Folder"). Fix inconsistencies.

1. Edge case coverage: Confirm each RFC's Edge Cases section is populated

(not empty or placeholder). Check that error/failure scenarios are addressed, not just happy paths.

1. Tech stack alignment: Verify code samples use the language, framework,

and libraries declared in the PRD's Technology Stack table.

1. Performance targets: Confirm specific numbers exist (not just "should be

fast") with acceptable ranges.

Present the review results to the user as a checklist with pass/fail for each item and details on anything that needs fixing.

Phase 6: Generate Implementation Prompt

After all RFCs are reviewed and finalized, ask the user:

"Would you like me to generate a complete implementation prompt for any of these RFCs? You can paste it into a new Claude Code session to implement the feature."

If the user says yes, ask which RFC(s) they want a prompt for (or offer to do all of them).

For large projects (3+ RFCs, parallel implementation): generate a full orchestration prompt instead of per-RFC prompts. An orchestration prompt assigns the agent an orchestrator role: it defines phases, spawns one sub-agent per RFC, verifies integration between phases, and never writes code directly. Use the orchestrator skill to execute this kind of prompt. Mention this option to the user when the implementation scope warrants it.

To generate the prompt, you must:

1. Read the project codebase — scan the actual project directory to identify:

• Existing files that will be modified or replaced
• Related services, utilities, or shared code the implementation will touch
• Project conventions (naming patterns, DI approach, architecture style)

1. Build the prompt with these sections in order:

a. Opening context — tell the agent what RFC to implement and where to find it (absolute path)

b. Files to read first — list every existing file the agent should read before writing code:

• The RFC document itself
• All files that will be modified/replaced
• Related RFCs for cross-cutting context
• Services or utilities the implementation depends on

c. Step-by-step implementation plan — ordered steps with:

• Specific file changes (what to create, modify, rename, delete)
• Detailed instructions for each change (not just "implement X" — spell out the logic)
• Dependencies between steps (what must happen before what)
• "Build after each major step to catch errors early"

d. Project conventions — any patterns the agent must follow:

• Architecture patterns (MVVM, Clean Architecture, etc.)
• DI approach, state management, naming conventions
• Custom colors, design tokens, shared components
• Testing patterns, deployment targets
• Feature flags or configuration to respect

1. Format the prompt as a single copyable block — the user should be able to paste it directly into a new Claude Code session with no edits needed.

What makes a good implementation prompt:

• Absolute file paths — no guessing, no relative paths from unknown roots
• Read before write — always tell the agent to read existing code first
• Ordered steps — numbered, with clear dependencies
• Specificity over brevity — spell out exactly what each step involves (which structs to rename, which enum cases to add, which UI sections to build)
• Explicit exclusions — "Do NOT touch X" or "That's handled in RFC-Y"
• Build verification — remind the agent to build/test after each major step
• Plan file — tell the agent to create a plan in tasks/todo.md first and mark items as it goes

See references/example-prompt-patterns.md for simpler prompt patterns that can be used as building blocks.

Adapting to project scale

Small project (landing page, simple tool, single-platform app):

• Combine PRD into a shorter format — skip personas if there's only one user type
• 2-3 RFCs may be sufficient
• Architecture overview can be a section within the PRD rather than a separate RFC

Medium project (multi-feature app, API + frontend):

• Full PRD with all sections
• 4-6 RFCs
• Separate architecture RFC (RFC-000)

Large project (cross-platform, multiple integrations, team of developers):

• Full PRD, possibly with appendices
• 6+ RFCs, potentially with sub-RFCs for complex areas
• Consider a "RFC-000: Architecture" that serves as the index document
• Add a dependency graph showing RFC implementation order

Reference files

• templates/prd-template.md — Full PRD template with section-by-section guidance
• templates/rfc-template.md — Full RFC template with section-by-section guidance
• references/job-stories-guide.md — How to write effective job stories
• references/example-prompt-patterns.md — Prompt patterns for feeding RFCs to coding agents