Refining Requirements

Requirements Refiner

Transform rough requirements into implementation-ready specifications optimized for Claude Code execution.

When to Use

• Received a PRD/spec that's prose-heavy and ambiguous
• Requirements scatter file paths throughout text
• Specs say "similar to X" without explaining X
• Missing API contracts, data models, or clear scope
• Before starting implementation on complex features
• Starting a new project from scratch (greenfield)
• Need to identify missing information before coding
• Architectural refactoring (e.g., session-based to JWT)
• Multi-service features spanning frontend/backend/database

When NOT to Use

Skip refinement for:

• Simple bug fixes with clear location (e.g., "fix typo in file X line Y")
• Well-structured Jira tickets that already have:
  • Explicit file paths to modify
  • Clear acceptance criteria
  • Defined scope and deliverables
• Single-line changes or trivial updates
• Tasks where implementation path is obvious

Decision rule: If the task has specific file paths, clear acceptance criteria, and no ambiguity about what to build → execute directly, don't refine.

Workflow

Phase 0: Auto-Detect Project Context

Step 1: Detect tech stack automatically

# Run these checks to auto-detect stack
ls package.json 2>/dev/null && echo "DETECTED: Node.js"
ls go.mod 2>/dev/null && echo "DETECTED: Go"
ls requirements.txt pyproject.toml 2>/dev/null && echo "DETECTED: Python"
ls pom.xml build.gradle 2>/dev/null && echo "DETECTED: Java"
ls Gemfile 2>/dev/null && echo "DETECTED: Ruby"
ls composer.json 2>/dev/null && echo "DETECTED: PHP"
ls Cargo.toml 2>/dev/null && echo "DETECTED: Rust"

Step 2: Read project conventions from CLAUDE.md

# Check for project conventions
cat CLAUDE.md 2>/dev/null || cat .claude/CLAUDE.md 2>/dev/null || echo "No CLAUDE.md found"

Extract and apply:

• Code style preferences
• Naming conventions
• Architecture patterns
• Testing requirements
• Documentation standards

Step 3: Detect project type

Greenfield indicators (any of these = greenfield):

• "Build new service/microservice"
• "From scratch" or "brand new"
• "No existing code"
• "Deploy to Kubernetes" (new service)
• No file paths to validate (all CREATE, no EXISTS)

Multi-stack indicators (any of these = multi-stack):

• Monorepo structure mentioned (packages/, services/)
• Multiple languages listed (React + Go + Python)
• Cross-service feature (frontend + backend + database)
• Shared types/contracts between services

Phase 1: Check for Missing Information

Scan PRD for completeness. See references/edge-cases.md for checklist.

If critical info is missing, note the gaps for the interview phase.

Phase 1.5: Interactive Deep-Dive Interview

This is the core differentiator. Don't just identify gaps - actively probe for them.

Step 1: Offer Skip Option

If PRD appears comprehensive (all sections present, clear scope, explicit file paths):

"Your requirements look comprehensive. Would you like to:
1. Proceed directly to spec generation
2. Quick sanity-check interview (5-10 questions)
3. Deep-dive interview (thorough exploration)"

If user chooses to proceed, skip to Phase 2.

Step 2: Conduct Interview

Use AskUserQuestion tool to probe deeply. Ask 1-4 questions per round.

Question Selection Priority:

1. Blocking unknowns (can't proceed without answer)
2. High-impact architectural decisions
3. Failure modes and edge cases
4. UX details and user segments
5. Operational concerns

Question Categories (see references/interview-patterns.md for full templates):

Question Quality Rules:

• Ask about WHAT HAPPENS WHEN, not WHAT IS
• Scenario-based > abstract
• Specific > general
• Probe edge cases, not happy paths
• Never ask obvious questions (database choice, endpoint names)

Example Good Questions:

"If the user closes the browser mid-transaction, what state should they see when they return?"

"When [Service A] and [Service B] both try to update the same record, which wins?"

"If this feature gets 100x expected traffic tomorrow, what breaks first?"

"What's the experience for a user on 3G mobile with 500ms latency?"

Example Bad Questions (NEVER ask these):

"What database will you use?" (obvious/they'll tell you)
"What should the API return?" (too low-level)
"Is security important?" (obviously yes)
"Should it be fast?" (obviously yes)

Step 3: Iterate Until Complete

WHILE interview_active:
    1. Review current knowledge and gaps
    2. Select 1-4 highest-priority questions
    3. Ask using AskUserQuestion with smart options
    4. Record answers
    5. Update mental model with new information
    6. Check exit conditions

Always include an exit option in each question set:

• "That's enough, proceed with spec" as an option
• Or ask explicitly: "Continue probing or ready to generate spec?"

Step 4: Exit Conditions

Stop interviewing when ANY of these are true:

• User explicitly says "proceed" or "that's enough"
• All critical gaps are filled (no blocking unknowns)
• Questions become repetitive or hypothetical
• Answers converge to "standard approach" or "same as before"

Anti-rationalization: If you're unsure whether to continue, ask one more round. Better to over-interview than under-interview.

Step 5: Summarize Findings

Before proceeding, briefly summarize key decisions from interview:

## Interview Summary

**Key Decisions:**
- Error handling: Show all errors at once, inline validation
- Scaling: Must handle 10k concurrent users by month 3
- Rollback: Feature flag with kill switch required
- Mobile: Responsive, not separate app

**Discovered Requirements:**
- Need audit trail for compliance
- Session timeout: 30 minutes idle
- Rate limit: 100 requests/minute per user

Phase 2: Extract and Classify

Pull out: scope, file mappings, data models, API contracts, business logic, UI copy

Incorporate interview findings:

• Add discovered requirements from interview
• Update scope based on tradeoff decisions
• Include edge cases and failure modes discussed
• Add compliance/security requirements surfaced
• Note any constraints or limitations mentioned

Phase 3: Validate File Paths Against Codebase

Actually verify paths exist:

# For each file path mentioned in PRD, check if it exists
find . -path "*/node_modules" -prune -o -name "filename.js" -type f -print

Read similar files to understand patterns:

# If PRD says "similar to existing controller", read it
cat app/controllers/v2/integrations/pos/pin/create.js 2>/dev/null | head -50

Mark paths as:

• ✓ EXISTS — file found, will modify
• ✗ CREATE — file doesn't exist, will create
• ? VERIFY — path mentioned but needs confirmation

Include validation summary in output:

## File Path Validation
| Path | Status | Action |
|------|--------|--------|
| app/models/egift.js | ✓ EXISTS | Modify — add field |
| app/controllers/v2/msite/verify.js | ✗ CREATE | New endpoint |

Phase 4: Structure

Apply template from references/templates.md

Adapt to detected tech stack using references/stack-patterns.md:

• File extensions and paths
• Migration format
• Naming conventions
• Test patterns

For multi-stack projects: Create SEPARATE file mapping sections per stack:

## File Mappings

### Frontend (React/TypeScript)
| Path | Status | Action |
|------|--------|--------|
| packages/web/src/components/Feature.tsx | ✗ CREATE | New component |

### Backend API (Go)
| Path | Status | Action |
|------|--------|--------|
| services/api/internal/handlers/feature.go | ✗ CREATE | New handler |

### Shared Types (TypeScript)
| Path | Status | Action |
|------|--------|--------|
| packages/shared/types/feature.ts | ✗ CREATE | Shared interfaces |

Include cross-service API contracts showing how services communicate.

Phase 5: Generate Test Scenarios

Derive tests from business logic. See references/test-patterns.md

Phase 6: Security & UI Copy

Security: Auto-generate checklist based on detected patterns. See references/security-checklist.md

UI Copy: Extract all user-facing text into bilingual table (ID/EN). See references/ui-copy-patterns.md

Phase 7: Verify Completeness

STOP before saving. Verify content and determine output format:

7a. Determine Output Format (Threshold-Based)

Estimate your total output line count using this heuristic:

Threshold: < 150 lines = single file, ≥ 150 lines = multi-file

Always use multi-file for (regardless of line count):

• Greenfield projects (creating a new service/package from scratch)
• Multi-stack features (different language ecosystems, e.g., React + Go)
• Features with 3+ API endpoints
• Features with external API integrations (webhooks, third-party services)
• Features with ANY TWO of: database migrations, API changes, UI copy

Single file is fine for:

• Simple bug fixes with clear scope
• Small features affecting 1-2 files
• Quick refinements that fit on one screen
• Client-side only changes (no backend)

Anti-rationalization rule: If you're trying to "fit" content into single-file, use multi-file instead. When uncertain, multi-file is always acceptable.

7b. Verify Section Content

For greenfield: Verify ALL file paths are marked CREATE (no EXISTS).

For multi-stack: Verify separate sections exist for each stack.

If any required section is missing: Complete it before proceeding.

Phase 8: Save Output

Use the format determined in Phase 7a (threshold-based).

Option A: Single File (< 150 lines or simple features)

mkdir -p agent-workflow/requirements

Naming: agent-workflow/requirements/YYYYMMDD-feature-name.md

Examples:

• agent-workflow/requirements/20241127-add-email-field.md
• agent-workflow/requirements/20241127-fix-validation-bug.md

Use the single-file template from Output Structure section below.

Option B: Multi-File Directory (≥ 150 lines or complex features)

mkdir -p agent-workflow/requirements/YYYYMMDD-feature-name

Directory structure:

agent-workflow/requirements/YYYYMMDD-feature-name/
├── README.md              # Overview + navigation (always)
├── scope.md               # In/Out scope (always)
├── file-mappings.md       # Files to create/modify (always)
├── api-contracts.md       # Endpoint definitions (if applicable)
├── data-model.md          # Database changes (if applicable)
├── business-logic.md      # Rules and flows (always)
├── test-scenarios.md      # Test cases (always)
├── security.md            # Security checklist (always)
├── ui-copy.md             # Bilingual UI text (if applicable)
└── execution-checklist.md # Implementation steps (always)

Examples:

• agent-workflow/requirements/20241127-whatsapp-verification/
• agent-workflow/requirements/20241127-new-verification-service/ (greenfield)

Conditional file creation (multi-file only):

• api-contracts.md — Only create if feature has API endpoints
• data-model.md — Only create if feature has database changes
• ui-copy.md — Only create if feature has user-facing text

Frontmatter (both formats):

---
title: [Feature Name]
created: [ISO timestamp]
status: draft | review | approved
stack: [detected stack]
type: feature | greenfield
original_prd: [filename if provided]
---

Phase 9: Offer Scaffolding (Optional)

After saving refined PRD, ask:

"Would you like me to scaffold the boilerplate files for this feature?"

If yes, create:

• Empty controller/handler files with correct structure
• Model files with field definitions
• Migration files with schema
• Test file skeletons
• Route registrations

Scaffolding rules:

• Read existing files to match code style
• Use project's linting/formatting conventions from CLAUDE.md
• Add TODO comments for implementation
• Don't overwrite existing files (create .new suffix if conflict)

Output Structure

For templates, see references/templates.md:

• Single-File Output Template — for simple features (< 150 lines)
• Multi-File Output Templates — for complex features (≥ 150 lines)

Multi-file directory structure:

agent-workflow/requirements/YYYYMMDD-feature-name/
├── README.md              # Overview + navigation
├── scope.md               # In/Out scope
├── file-mappings.md       # Files to create/modify
├── api-contracts.md       # (if applicable)
├── data-model.md          # (if applicable)
├── business-logic.md      # Rules and flows
├── test-scenarios.md      # Test cases
├── security.md            # Security checklist
├── ui-copy.md             # (if applicable)
└── execution-checklist.md # Implementation steps

Claude Code Commands

Quick reference for commands used by this skill:

# Auto-detect stack
ls package.json go.mod requirements.txt pyproject.toml pom.xml 2>/dev/null

# Read project conventions
cat CLAUDE.md 2>/dev/null

# Validate file exists
test -f "path/to/file.js" && echo "EXISTS" || echo "CREATE"

# Read existing pattern
cat path/to/similar/file.js | head -100

# Option A: Single file (< 150 lines)
mkdir -p agent-workflow/requirements
cat > agent-workflow/requirements/YYYYMMDD-feature-name.md << 'EOF'
[all sections in one file]
EOF

# Option B: Multi-file directory (≥ 150 lines or complex)
mkdir -p agent-workflow/requirements/YYYYMMDD-feature-name
cat > agent-workflow/requirements/YYYYMMDD-feature-name/README.md << 'EOF'
[overview content]
EOF
cat > agent-workflow/requirements/YYYYMMDD-feature-name/scope.md << 'EOF'
[scope content]
EOF
# ... repeat for each required file

# Scaffold files (with conflict check)
test -f "path/to/file.js" && cp "path/to/file.js" "path/to/file.js.new" || touch "path/to/file.js"

Examples

See references/transformation-patterns.md for detailed before/after examples including:

• API integration features
• Data model changes
• Complex verification flows

Anti-Patterns

1. Narrative flow — Use structured sections, not stories
2. Embedded paths — Consolidate all paths in File Mappings
3. Implicit comparisons — "Like existing flow" needs explanation
4. Mixed concerns — Separate UI copy from business logic
5. Missing edge cases — Explicitly list: what if X fails?
6. Not reading existing code — Always cat similar files first
7. Not saving output — Always save to agent-workflow/requirements/
8. Over-refining structured specs — Don't bloat already-clear Jira tickets
9. Mixed stack patterns — Don't use Node.js patterns for Go projects
10. Missing completeness check — Always verify all sections before saving