Refinement Process

Quick Reference

The refinement process transforms Linear features from Discovery into comprehensive mini-PRDs with Behaviour scenarios, visual designs, and complete technical specifications. This is the critical bridge between understanding what to build (Discovery) and building it (Delivery).

Typical duration: 2-6 hours per feature End result: Mini-PRD, Markdown scenarios, Storybook stories, RTL tests, Selector contract

When to Use

✅ Use Refinement Process for:

• Transforming Linear features into detailed specifications
• Creating Behaviour scenarios from acceptance criteria
• Defining technical implementation approaches
• Preparing features for Delivery phase

❌ Not Refinement phase:

• Product-level exploration (use Discovery)
• Implementation work (use Delivery)
• Bug fixes without new behavior (go directly to Delivery)

Git Workflow

Refinement work happens on the specs branch (see ADR-009).

# Start refinement on a specific issue
start-refine myproject LIN-123  # Creates/attaches to specs worktree + tmux

The specs branch is shared by PM, UX, and QA for all pre-implementation work. During Refinement:

When Refinement complete (before approval):

1. Merge specs into develop (direct merge, no PR — documents don't need CI)
2. THEN approve the issue in Linear (status → "Approved")
3. "Approved" means develop is ready — specs are already there
4. SE creates issue branch via start-work (branches from develop with specs included)

Trunk-Based Projects

Projects configured with branching.strategy: trunk-based work directly on the main branch instead of a specs branch. For these projects:

• start-refine uses the main working directory (no specs worktree)
• All Refinement artifacts go directly to the trunk branch
• The workflow is otherwise identical

See Work Management Guide for full branching details.

Core Principles

What is a Feature?

A Feature is an atomic unit of work - one thing, front-to-back:

Each Feature gets:

• ONE Linear Issue
• ONE Mini-PRD
• ONE E2E test file
• All its scenarios (happy path + errors + edge cases)

See Atomic Refinement Guide for outside-in development and full hierarchy.

1. Three Amigos Analysis

Every feature MUST be analyzed from three perspectives:

Business Perspective:

• What value does this deliver?
• What are the success criteria?
• How does this impact users?

Development Perspective:

• What technical approach should we use?
• What are the dependencies?
• What could be reused or extended?

QA Perspective:

• What could go wrong?
• What edge cases exist?
• How do we test this?

2. Human-in-the-Loop

There are two mandatory approval gates:

Gate 1 - Initial Approach (after Three Amigos):

• ✋ Technical approach MUST be validated
• ✋ Scope must be confirmed
• ✋ Prevents expensive wrong-direction work

Gate 2 - Complete Specifications (before Delivery):

• ✋ Scenario coverage MUST be confirmed
• ✋ Visual designs MUST be approved
• ✋ All artifacts MUST have explicit sign-off

NEVER proceed to Delivery without Gate 2 approval

3. Comprehensive Specification

The mini-PRD MUST contain everything needed for implementation:

• Business context and success criteria
• Technical architecture and API contracts
• Behaviour scenarios with complete coverage (happy/error/edge)
• Visual designs as executable Storybook stories
• RTL tests locking component behavior
• Selector contract bridging design and tests
• Seed data requirements for testing
• Implementation notes (dependencies, performance, security)

Workflow Overview

Step 0: Scope & Context Assessment

Before diving into analysis, assess scope and project state:

0a. Epic Detection (MANDATORY)

Evaluate if this is an Epic that needs delivery chunking:

If Epic detected:

1. Label as Epic in Linear (if not already)
2. Identify Delivery Chunks (atomic units that can be implemented/deployed independently)
3. Proceed with ONE PRD that includes a "Delivery Chunks" section
4. After approval, create child issues for each chunk

Entity Epics from Discovery: If this issue came from entity-driven Discovery, it may already be labeled as an Epic with capability sub-issues. Check the description for "This is an entity Epic" — if present, the delivery chunks are the capability sub-issues. Write one PRD covering all capabilities; each capability maps to a delivery chunk.

Delivery Chunk Criteria:

• Can be implemented in one PR
• Has a clear "done" state
• Can be deployed independently (given dependencies)
• Doesn't leave the system in a broken state

Note: One PRD per Epic is fine. Chunking is for DELIVERY, not requirements.

0b. Project Context

Greenfield Advantage: When there are no live users, prefer:

• Correct naming from the start (don't defer renames)
• Clean architecture (don't accumulate tech debt)
• Full schema design (not incremental patches)

Brownfield Caution: When users exist, consider:

• Migration complexity and rollback plans
• Backwards compatibility requirements
• Phased rollout strategies

Step 1: Start Refinement

Command: /refinement:refine LIN-123

Orchestrator actions:

1. Read Linear feature and acceptance criteria
2. Read linked Discovery documentation
3. Assess project context (greenfield/brownfield, terminology changes)
4. Analyze from Three Amigos perspectives
5. Present initial analysis for human feedback

Approval Gate #1: User must approve initial approach before proceeding to detailed specs

After approach approval — Initial Estimation: Once the approach is approved, load af-estimate-effort and produce an initial refined estimate:

• Decompose by functional area (FE, BE, INFRA, TEST, UX, DOCS)
• Estimate hours per area, convert to fibonacci story points
• Confidence level: typically Medium (approach known, details to come)
• Tag as [Refined Estimate] in Linear comment
• Update Linear issue with story points

This estimate will be revisited after agents complete detailed specs (Step 3b).

Step 1a: Create Specification Sub-issues

For each parent issue entering Refinement, create sub-issues for specification work.

[Behaviour] and [UX] sub-issues are exempt from the sub-issue size policy — they serve a structural BDD purpose and are always created regardless of parent size.

[Behaviour] Sub-issue:

Title: [Behaviour] {Parent Title} scenarios
Label: type:behaviour
Assignee: QA role holder (or whoever wears the QA hat — see project CLAUDE.md)
Cycle: Current refinement cycle
Estimate: Typically 1-3 points (part of parent total)
Status: Approved

[UX] Sub-issue (if feature has UI):

Title: [UX] {Parent Title} Storybook
Label: type:ux
Assignee: UX role holder (or whoever wears the UX hat — see project CLAUDE.md)
Cycle: Current refinement cycle
Estimate: Typically 1-3 points (part of parent total)
Status: Approved

Ordering: Create and complete the [Behaviour] sub-issue first. Behaviour scenarios define WHAT the feature does; UX designs define HOW it looks. UX work uses the Behaviour scenarios as input.

Other sub-issues (non-spec): Only create additional sub-issues when the parent exceeds 2 days (12 hours) of human effort, each sub-issue is ≥2 days, and the work is genuinely independent (different person, different phase, or independently deliverable). Otherwise, use a checklist in the parent issue description. See af-manage-work-state for the full sub-issue creation policy.

After creating sub-issues, apply approval labels to the parent issue:

• Apply approval:bdd-pending label to parent when [Behaviour] sub-issue is created
• Apply approval:ux-pending label to parent when [UX] sub-issue is created (UI features only)

Linear structure after sub-issue creation:

Issue: Login Flow (8 pts) [Refining]
  ├── Sub-issue: [Behaviour] Login Flow scenarios (2 pts) [Approved] @QA, Cycle 12
  ├── Sub-issue: [UX] Login Flow Storybook (2 pts) [Approved] @UX, Cycle 12
  └── Description contains implementation checklist (not sub-issues)

Parent estimate includes sub-issue effort: The parent's 8 points includes the 2+2 for specs plus ~4 for implementation. Sub-issue estimates are a breakdown, not additions.

No [API] sub-issue: API contract work happens early in Delivery as part of implementation. It's a checklist item in the parent issue, not a separate spec artifact requiring sign-off.

Branch for Spec Work

Feature branch model:

• Work on the specs branch (via start-refine command)
• Commit Behaviour scenarios and Storybook stories to this branch
• All refinement work for a cycle shares the same specs worktree
• PR to develop when all spec sub-issues for that cycle are approved

Trunk-based model:

• Work directly on the default branch
• Commit spec artifacts alongside other changes
• No separate specs branch

Sub-issues do not get their own branches: All spec work for a refinement period goes to the same branch. The branch corresponds to the work phase, not the individual sub-issue.

Step 2: Create Mini-PRD

Once approach approved, orchestrator creates mini-PRD with 8 sections:

Section 1: Feature Summary

• Purpose, description, success criteria
• User value proposition
• Business context

Section 2: Discovery Context

• Links to user research
• Related ADRs (architectural decisions)
• Technical constraints from Discovery

Section 3: Technical Specification

• API contracts (REST, GraphQL, gRPC, etc.)
• Database models (entities, relationships, migrations)
• Architecture diagrams (system components, data flow)
• Integration points (external services, APIs)

Section 4: Behaviour Scenarios

• Markdown scenario specifications (Preconditions/Steps/Expected format)
• Test type classification (E2E, Integration, Component, Unit)
• Happy path scenarios
• Error handling scenarios
• Edge case scenarios
• Selector contract references for UI scenarios

Section 5: Visual Specifications

• Storybook stories for all UI states
• Selector contract (shared test IDs)
• RTL tests (component behavior contracts)
• Component specifications
• Responsive behavior
• Accessibility requirements

Section 6: Implementation Notes

• Dependencies and prerequisites
• Seed data requirements
• Performance targets
• Security considerations
• Monitoring requirements
• Infrastructure testing requirements (see below)

Infrastructure Testing Requirements

When a feature requires new infrastructure code, specify testing requirements:

Utility Modules (lib/ or similar):

API Routes:

Rule: Any new utility module or API route MUST have corresponding test requirements in the mini-PRD. Behaviour scenarios test user behaviour; infrastructure tests ensure the plumbing works.

Section 7: Effort Estimation

• Story points and total hours
• Breakdown by functional area (hours and role)
• Sub-issue estimates with hours and points roll-up (for features > 3 points)
• Confidence level and estimation basis
• Change from Discovery estimate (if applicable)

Section 8: Delivery Chunks (for Epics)

Include this section when the feature is an Epic with multiple delivery chunks.

For each chunk, specify:

• What's included (components, APIs, infrastructure)
• What it depends on (other chunks, external blockers)
• Which scenarios it covers (by number)
• Expected PR count (usually 1)

After approval: Create Linear child issues for each chunk.

Section 8: Delivery Checklist

• Pre-implementation checks
• During implementation milestones
• Post-deployment verification

Step 3: Invoke Specialized Agents

Orchestrator delegates specialized work:

Behaviour Agent Invocation:

Task → bdd-agent
Input: Mini-PRD content, Linear issue ID, glossary path
Output:
  - Markdown scenarios in mini-PRD Section 4
  - /tests/selectors/[capability].ts (selector contract)
Purpose: Transform requirements into scenario specifications

UX Design Agent Invocation (for UI features):

Task → ux-design-agent
Input: Behaviour scenarios, component requirements, selector contract
Output:
  - /stories/[capability]/[feature-name].stories.tsx
  - /src/components/[feature]/[Component].test.tsx (RTL tests)
  - /docs/requirements/[feature-name]/user-flow.md
Purpose: Create visual specifications and lock behavior with RTL tests

Architecture Agent Invocation (for complex features):

Task → architecture-quality-agent
Input: Technical specification, existing architecture
Output: Architecture review, ADR updates
Purpose: Validate technical approach against standards

After agent completion, update approval labels on parent:

• bdd-agent success → Replace approval:bdd-pending with approval:bdd-approved on parent
• ux-design-agent success → Replace approval:ux-pending with approval:ux-approved on parent

Step 3b: Agent Invocation Checklist

MANDATORY CHECKPOINT - Before proceeding to Step 4, verify ALL applicable items:

How to check "Does feature have UI components?"

• Will users see new screens, forms, or interactive elements? → YES
• Are there new components to build (buttons, cards, modals)? → YES
• Does the mini-PRD Section 5 list visual specifications? → YES
• Is this pure backend/infrastructure with no UI? → NO

Do NOT proceed to Step 4 until this checklist is complete.

If you skipped an agent, STOP and invoke it now. It's cheaper to catch this here than after approval.

Step 4: Quality Validation

Completeness Checks:

• ✅ All Linear acceptance criteria addressed
• ✅ Discovery context properly linked
• ✅ Technical approach documented
• ✅ All scenarios covered (happy/error/edge)

Quality Checks:

• ✅ Glossary compliance verified (scenario terms)
• ✅ Selector contract exists for UI features
• ✅ Storybook play functions test UI component behaviors (primary)
• ✅ RTL tests cover non-visual logic (hooks, utils, state management)
• ✅ No duplicated assertions between Storybook and RTL
• ✅ Design system consistency (shadcn/ui components)
• ✅ Accessibility requirements specified
• ✅ Performance targets defined
• ✅ Test type classified for each scenario

Estimation Checks:

• ✅ Effort estimates finalized (story points and hours set on Linear issue)
• ✅ Mini-PRD Section 7 (Effort Estimation) complete with breakdown
• ✅ For features > 3 points: sub-issues created with individual estimates
• ✅ Hours AND points roll up to parent feature (not just Linear's points roll-up)
• ✅ Estimate tagged as [Refined Estimate] in Linear comment

Artifact Checks:

• ✅ Mini-PRD document created with all 8 sections
• ✅ Scenarios use Markdown format (not Gherkin)
• ✅ Selector contract defined (/tests/selectors/)
• ✅ Storybook stories with play functions created and passing (for UI - primary component tests)
• ✅ RTL tests created and passing for non-visual logic (hooks, utils, state management)
• ✅ User flows documented
• ✅ Glossary updated (if new terminology introduced)
• ✅ Ripple effects assessed (if terminology changes)
• ✅ UX ripple effects documented (even for infrastructure features)
• ✅ Affected story files listed explicitly (if terminology changes)

Step 5: Commit and Validate

IMPORTANT: Commit artifacts BEFORE requesting approval. The git commit hook acts as a quality gate.

Process:

1. Stage all requirement artifacts (mini-PRD, scenarios, selectors, ADRs)
2. Attempt commit - hook will fire with checklist
3. Work through hook checklist honestly:
   • TESTS: Storybook play functions + RTL for non-visual logic should PASS (components exist). E2E/Integration use test.todo() (implemented in Delivery). Never test.skip() — it hides gaps silently.
   • DOCUMENTATION: YES - this IS the documentation
   • VALIDATION: Run docs-quality-agent if prompted
   • ARCHITECTURE: Confirm ADRs created for decisions made
   • SECURITY: Verify no secrets in docs
4. Complete commit after addressing all items
5. Only THEN proceed to human approval

Why commit first:

• Hook catches missing ADRs, validation, and other issues
• Issues found before asking user to approve
• Avoids "approved but incomplete" state

Step 6: Human Approval

Approval Gate #2: Present complete mini-PRD to user for review

Command: /refinement:approve LIN-123

Pre-conditions for approval:

• All [Behaviour] sub-issues are Done
• All [UX] sub-issues are Done (if feature has UI)
• Parent issue has approval:bdd-approved label
• Parent issue has approval:ux-approved label (if UI feature)
• Mini-PRD is complete and committed
• Git commit hook checklist passed
• Spec artifacts merged to develop (direct merge before approval)

Approval process:

1. Verify commit completed (Step 5)
2. Verify all sub-issues are Done
3. Verify all artifacts exist
4. Run quality checks (docs-quality-agent) if not done in Step 5
5. Merge specs → develop (direct merge, no PR — documents don't need CI)
6. Get explicit human approval
7. Update Linear parent issue to "Approved"
8. "Approved" means develop is ready — specs are already there
9. Parent issue ready for scheduling into delivery cycle

Step 7: Post-Approval - Create Child Issues (for Epics)

If this is an Epic with Delivery Chunks (see Step 0a):

After approval, create child issues in Linear for each chunk (via GraphQL API - see af-query-linear-api):

# Example: Create child issue with parent reference
curl -s -X POST https://api.linear.app/graphql \
  -H "Content-Type: application/json" \
  -H "Authorization: $LINEAR_API_KEY" \
  -d '{
    "query": "mutation { issueCreate(input: { title: \"Chunk 1: Photo Capture Flow\", teamId: \"TEAM-UUID\", parentId: \"EPIC-UUID\", description: \"Refs PRD Section 7 Chunk 1. Scenarios 1-18.\" }) { success issue { identifier } } }"
  }'

For each child issue:

1. Reference the parent PRD and chunk section
2. List which scenarios it covers
3. Set "blocked by" if depends on other chunks
4. Link to relevant components/stories

Linear structure after approval:

EPIC-123: Item Intake & Photo Capture [epic] [Approved]
  ├── EPIC-123-1: Photo Capture Flow [Approved]
  ├── EPIC-123-2: Draft Recovery [Approved] [blocked by EPIC-123-1]
  ├── EPIC-123-3: S3 Upload [Approved] [blocked by EPIC-123-1]
  └── EPIC-123-4: Lambda Thumbnails [Approved] [blocked by EPIC-123-3]

Do NOT move to Delivery without approval

Decision Points

Does this feature need UX Design Agent?

• Has UI components? → YES (invoke ux-design-agent for Storybook + RTL)
• API/backend only? → NO (skip visual specs and RTL)
• Modifies existing UI? → YES (update Storybook stories and RTL tests)

How many scenarios should I create?

• Minimum: 1 happy path + 1 error case
• Typical: 4-6 (happy, errors, edge cases)
• Complex features: 8-15 (multiple paths, states, conditions)
• Rule: Cover all acceptance criteria + discovered edge cases

Which test type for each scenario?

• E2E (~10%): Complete user journeys through real UI + real backend
• Integration (~20%): API/backend behavior without UI
• Component (~20%): UI behavior with RTL (forms, interactions, accessibility)
• Unit (~50%): Pure validation logic (email format, calculations)
• Rule: Push tests down the pyramid - use minimum scope needed

Which agents do I need?

• Always: bdd-agent (for Behaviour scenario specifications)
• UI features: ux-design-agent (for Storybook + RTL tests)
• Complex architecture: architecture-quality-agent (for review)
• Final validation: docs-quality-agent (for documentation standards)

Does this feature involve terminology changes?

When renaming entities, attributes, or roles (e.g., Tenant → Organisation), assess the full ripple effect:

Greenfield rule: Do all renames now while it's free. Brownfield rule: Create migration plan with backwards compatibility.

Does this feature require new ADRs?

IMPORTANT: Technical decisions made during Refinement phase may warrant ADRs. This is especially true when you ask the user to choose between approaches during Three Amigos analysis.

Decision Framework (from af-decide-architecture):

Rule: If you presented options to the user during Three Amigos (e.g., "client-side only vs server-side validation"), that decision likely warrants an ADR.

Process:

1. When a technical decision is made, check against the framework above
2. If ADR warranted, create it BEFORE finalizing mini-PRD
3. Reference the ADR in mini-PRD Section 2 (Discovery Context)
4. ADR must be committed with mini-PRD

Common decisions that need ADRs:

• Validation strategy (client vs server vs both)
• Storage approach (localStorage vs server-side vs IndexedDB)
• API design patterns (REST vs GraphQL, sync vs async)
• Authentication flows
• Data model design

→ Load af-decide-architecture for full ADR workflow and templates

Does this feature have UX ripple effects?

IMPORTANT: Even infrastructure features may require UX changes. Always check:

1. Search for affected stories:

   bashCopy

   grep -r "OldTerm\|oldTerm" stories/
   

2. For each affected story, document in mini-PRD Section 5:

   • File path
   • Specific changes required (renames, mock data, props)
   • New stories needed (if any)
   • data-testid additions required

3. Add Component-level Behaviour scenarios for any UI behavior changes

4. Update Delivery Checklist with specific UX tasks

Common UX ripple effects:

Rule: If terminology or auth changes, there ARE UX changes. List them explicitly.

Common Patterns

API-Heavy Features

Focus on:

• API contract definitions (REST/GraphQL/gRPC schemas)
• Request/response examples with payloads
• Error codes, messages, and handling
• Rate limiting and throttling considerations
• Authentication and authorization requirements
• Integration test scenarios

Example: User authentication, payment processing, data exports

UI-Heavy Features

Focus on:

• Storybook stories (all states: default, loading, error, empty, success)
• Selector contract (shared data-testid values)
• RTL tests (component behavior contracts)
• Responsive breakpoints
• Interaction animations
• Accessibility (ARIA labels, keyboard nav)

Example: Dashboard widgets, forms, data visualization

Data-Heavy Features

Focus on:

• Database schema definition (SQL/NoSQL/Graph)
• Data migration plans and rollback strategies
• Seed data requirements for testing
• Backup/recovery procedures
• Data validation rules and constraints
• Query performance and indexing strategy

Example: Reporting systems, bulk operations, data imports

Infrastructure Features

Infrastructure features are foundational work that enables other features but isn't directly user-facing. Examples: schema extensions, auth configuration, new data models.

Focus on:

• Authorization rules and data isolation
• API contracts (GraphQL schema, Lambda signatures)
• Data integrity and validation
• Integration test scenarios (not E2E user journeys)
• Seed data for dependent features
• Glossary updates for new entities

Scenarios emphasize:

• CRUD operations with correct authorization
• Tenant/shop isolation (cross-boundary rejection)
• Role-based access control
• Data validation and error handling
• Edge cases (empty states, max limits, invalid input)

Scenarios de-emphasize:

• Full UI user journeys
• New Storybook stories (unless admin UI exists)

BUT STILL CHECK for UX ripple effects:

• Terminology changes → existing stories need updating
• Auth changes → login/signup forms need updating
• Role changes → badges and dropdowns need updating

Always run: grep -r "OldTerm" stories/ to find affected files.

Example: Schema migrations, new data models, Cognito configuration, Lambda handlers

File Outputs

Refinement phase Creates (Specifications)

.claude/templates/
  └── glossary.yml          # UPDATE: Add new terms, deprecate old

docs/requirements/[feature-name]/
  ├── mini-prd.md           # Complete specification
  │   ├── Section 4: Scenarios (Markdown format)
  │   └── Section 5: Selector contract references
  └── user-flow.md          # Mermaid flow diagrams

tests/selectors/
  └── [capability].ts       # Selector contract (shared test IDs)

stories/
  ├── [Component].stories.tsx       # Shared component stories
  └── [capability]/                 # Feature-specific stories
      └── [Feature].stories.tsx

src/components/[feature]/
  └── [Component].test.tsx  # RTL tests (colocated with components)

Glossary Updates (When Introducing Terminology)

When a feature introduces new entities, roles, or statuses, update the glossary:

# .claude/templates/glossary.yml

# Add new terms
Organisation:
  definition: "A charity or business using the platform (formerly Tenant)"
  replaces: Tenant  # Mark what it replaces

Shop:
  definition: "A physical location within an Organisation"

org_admin:
  definition: "Organisation-wide administrator role"
  replaces: owner

# Deprecate old terms
Tenant:
  deprecated: true
  replaced_by: Organisation
  removal_version: "2.0"

Selector Contracts (Bridge Between Stories/RTL/E2E)

Selector contracts are the single source of truth for data-testid values:

// tests/selectors/organisation.ts
export const organisationSelectors = {
  // Form fields
  nameInput: 'organisation-name-input',
  logoUpload: 'organisation-logo-upload',
  submitButton: 'organisation-submit-button',

  // Display elements
  nameDisplay: 'organisation-name-display',
  statusBadge: 'organisation-status-badge',
} as const;

Usage contract:

• Storybook stories → data-testid={organisationSelectors.nameInput}
• RTL tests → screen.getByTestId(organisationSelectors.nameInput)
• E2E tests → page.getByTestId(organisationSelectors.nameInput)

This ensures refactoring a selector updates all consumers.

Delivery Phase Creates (Executable Tests)

tests/
  ├── e2e/
  │   └── [feature].spec.ts        # Playwright E2E specs
  ├── integration/
  │   └── [feature].test.ts        # Jest integration tests
  ├── unit/
  │   └── [module].test.ts         # Jest unit tests
  ├── selectors/                   # From Refinement
  ├── fixtures/                    # Test data
  └── helpers/                     # Shared test utilities

amplify/seed/
  └── [feature].ts          # Seed data for testing (if needed)

Key Principle

• Refinement creates scenarios (specifications) in mini-PRD
• Delivery generates executable test files FROM those specifications
• Selector contract bridges both phases (shared source of truth)

Seed Data Requirements

When scenarios have preconditions like "Given a Super Admin account exists", Refinement phase must specify seed data:

In mini-PRD Section 6 (Implementation Notes):

### Seed Data Requirements

**Users:**
| Email | Password | Type | Attributes |
|-------|----------|------|------------|
| [email protected] | SuperAdmin123! | Platform | platform_role: super_admin |
| [email protected] | OrgAdmin123! | Organisation | role: admin, tenant_id: {seed} |

**Organisations:**
| Name | Tenant ID | Status |
|------|-----------|--------|
| Test Org | auto-generated | active |

**Location:** `/amplify/seed/[feature].ts`
**Command:** `npm run seed`

Seed data is idempotent - checks existence before creating.

Common Pitfalls

1. Skipping Three Amigos analysis

   • All three perspectives MUST be considered
   • Missing perspective leads to incomplete specs

2. Proceeding without human approval

   • NEVER skip either approval gate
   • Gate 1 prevents wrong-direction work
   • Gate 2 prevents incomplete implementation

3. Incomplete scenario coverage

   • Must include happy path, errors, edge cases
   • Missing scenarios = bugs in production

4. Missing selector contract

   • UI scenarios MUST reference selectors
   • Components MUST use data-testid from contract
   • Tests MUST import selectors

5. Missing RTL tests

   • Refinement phase creates RTL tests
   • Engineering keeps them passing during implementation
   • RTL tests are the "design contract"

6. Generic specifications

   • Mini-PRDs must be specific and detailed
   • Generic specs lead to implementation ambiguity

7. Missing technical details

   • API contracts MUST be complete
   • Database schemas MUST be defined
   • Missing details cause delays in Delivery

8. Ignoring glossary

   • Behaviour scenarios MUST use glossary terms
   • Inconsistent terminology confuses stakeholders

9. All scenarios as E2E

   • Push tests down the pyramid
   • E2E only for critical complete journeys
   • Most scenarios should be Component or Unit

10. Deferring renames in greenfield projects

    • No live users = no migration risk
    • "Do it right now" beats "fix it later"
    • Deferred renames accumulate as tech debt

11. Forgetting glossary updates

    • New entities MUST be added to glossary
    • Old terms MUST be marked deprecated
    • Inconsistent terminology causes confusion across teams

12. Missing ripple effect assessment

    • Terminology changes touch many layers
    • Schema → handlers → auth → UI → stories → tests → docs
    • Incomplete assessment leads to broken references

13. Assuming infrastructure features have no UX impact

    • Entity renames affect story files and display text
    • Auth changes affect login/signup forms
    • Role changes affect badges, dropdowns, permissions UI
    • Always grep stories/ for affected terms
    • List specific story file changes in Section 5

14. Forgetting ADRs for decisions made during requirements

    • If you asked the user to choose between approaches → probably needs ADR
    • Technical decisions made in Three Amigos should be evaluated
    • Create ADR BEFORE finalizing mini-PRD, not after
    • Reference ADR in mini-PRD Section 2

15. Requesting approval before committing

    • Always commit artifacts before asking for human approval
    • Git commit hook acts as quality gate (tests, docs, architecture, security)
    • Hook catches issues that should be fixed before approval
    • Avoids approving incomplete specifications

16. Skipping UX Design Agent for UI features

    • If feature has UI components → MUST invoke ux-design-agent
    • Creates Storybook stories (visual specifications)
    • Creates RTL tests (component behavior contracts)
    • Missing stories = incomplete specifications
    • Use Step 3b checklist to catch this before proceeding

17. Missing Epic detection and delivery chunking

    • Large features with multiple capabilities should be recognized as Epics
    • One PRD is fine, but must include Delivery Chunks section
    • Without chunks, Engineering faces overwhelming context
    • After approval, create child issues for each chunk
    • See Step 0a for Epic detection signals

Integration with Other Phases

Before Refinement:

• ← Discovery Phase - Provides Linear features and project context

After Refinement:

• → Delivery Phase - Implements mini-PRD specifications
• → Delivery uses scenarios as test specifications
• → Delivery keeps RTL tests passing
• → Delivery writes E2E tests from scenario specs

Refinement provides:

• Complete implementation blueprint
• Scenario specifications for test generation
• Selector contract for consistent testing
• RTL tests as component behavior contracts
• Approval to begin development
• Effort estimates (story points where 1 point = half day)

Detailed Reference

For complete requirements documentation:

• Read .claude/docs/guides/refinement-guide.md (full workflow with examples)

For scenario creation:

• Load af-write-bdd-scenarios skill
• Read .claude/docs/guides/testing-guide.md

For mini-PRD template:

• Read .claude/templates/mini-prd-template.md

For agent interactions:

• Read .claude/agents/bdd-agent.md
• Read .claude/agents/ux-design-agent.md
• Read .claude/agents/af-work-management-agent.md

Success Criteria

Refinement complete when:

• ✅ Scope assessed - Epic detected and labeled if applicable (Step 0a)
• ✅ Project context assessed (greenfield/brownfield)
• ✅ Three Amigos analysis performed and approved (Gate 1)
• ✅ Complete mini-PRD created with all sections (7 for features, 8 for epics)
• ✅ Delivery Chunks documented (for Epics - Section 7)
• ✅ Markdown scenarios cover all acceptance criteria + edge cases
• ✅ Selector contract defined for UI features
• ✅ Storybook stories with play functions created and passing (for UI features - primary component tests)
• ✅ RTL tests created and passing for non-visual logic (hooks, utils, state management)
• ✅ Seed data requirements documented
• ✅ Glossary updated (if new terminology)
• ✅ Ripple effects documented (if terminology changes)
• ✅ ADRs created for architectural decisions made during requirements
• ✅ All artifacts committed (git commit hook checklist passed)
• ✅ All artifacts validated by quality agents
• ✅ Human has explicitly approved specifications (Gate 2)
• ✅ Linear issue moved to "Approved"
• ✅ Child issues created for each Delivery Chunk (for Epics - Step 7)