product-expert-design

Product Expert Design

Guide for designing agent experts that serve end users through adaptive, personalized experiences.

Codebase Experts vs Product Experts

Aspect	Codebase Expert	Product Expert
Scope	One per domain	One per user
Storage	File system (YAML)	Database (JSONB)
Updates	After code changes	After user actions
Size	300-1000 lines	Typically smaller
Latency	Not critical	Must be fast
Privacy	Internal only	User data concerns

When to Use

• Building product features that learn from user behavior
• Creating per-user expertise files for personalization
• Implementing AI-driven adaptive UX
• Designing recommendation systems with user mental models
• Evaluating whether product experts are appropriate for your use case
• Building progressive personalization with latency considerations

Product Expert Architecture

┌─────────────────────────────────────────────────────┐
│ User Action (view, click, purchase, etc.)           │
└──────────────────────┬──────────────────────────────┘
                       │
                       ▼
┌─────────────────────────────────────────────────────┐
│ Action Tracker                                      │
│ • Capture action type                               │
│ • Record context (time, device, etc.)               │
│ • Queue for expertise update                        │
└──────────────────────┬──────────────────────────────┘
                       │
                       ▼
┌─────────────────────────────────────────────────────┐
│ User Expertise Store (Database)                     │
│ • Per-user JSONB column                             │
│ • Structured preference model                       │
│ • Behavior patterns                                 │
└──────────────────────┬──────────────────────────────┘
                       │
                       ▼
┌─────────────────────────────────────────────────────┐
│ UI Generation Agent                                 │
│ • Load user expertise                               │
│ • Generate personalized UI                          │
│ • Adapt recommendations                             │
└─────────────────────────────────────────────────────┘

User Expertise Schema

{
  "user_id": "uuid",
  "created_at": "timestamp",
  "updated_at": "timestamp",

  "preferences": {
    "categories": ["tech", "sports"],
    "price_range": {"min": 50, "max": 500},
    "brands": ["Apple", "Sony"],
    "style": "minimalist"
  },

  "behavior_patterns": {
    "active_hours": [9, 10, 11, 14, 15, 20, 21],
    "device_preference": "mobile",
    "session_length_avg": 420,
    "purchase_frequency": "monthly"
  },

  "interaction_history": {
    "views": [
      {"item_id": "123", "timestamp": "...", "duration": 45}
    ],
    "cart_adds": [
      {"item_id": "456", "timestamp": "..."}
    ],
    "purchases": [
      {"item_id": "789", "timestamp": "...", "amount": 299}
    ]
  },

  "inferred_interests": {
    "high": ["wireless headphones", "smart home"],
    "medium": ["fitness trackers"],
    "low": ["gaming"]
  },

  "recommendations_context": {
    "last_shown": ["item1", "item2"],
    "clicked": ["item1"],
    "dismissed": ["item3"]
  }
}

Act-Learn-Reuse for Products

ACT: User Takes Action

async function trackUserAction(
  userId: string,
  action: UserAction
): Promise<void> {
  // Record the action
  await db.userActions.create({
    userId,
    actionType: action.type,
    context: action.context,
    timestamp: new Date()
  });

  // Queue expertise update
  await expertiseQueue.add({
    userId,
    action,
    priority: getActionPriority(action.type)
  });
}

LEARN: Update User Expertise

async function updateUserExpertise(
  userId: string,
  action: UserAction
): Promise<void> {
  // Load current expertise
  const expertise = await loadUserExpertise(userId);

  // Update based on action type
  switch (action.type) {
    case 'view':
      updateViewPatterns(expertise, action);
      break;
    case 'cart_add':
      updatePurchaseIntent(expertise, action);
      break;
    case 'purchase':
      updatePreferences(expertise, action);
      break;
  }

  // Recalculate inferred interests
  expertise.inferred_interests = inferInterests(expertise);

  // Save updated expertise
  await saveUserExpertise(userId, expertise);
}

REUSE: Personalized Experience

async function generatePersonalizedUI(
  userId: string
): Promise<UIConfig> {
  // Load user expertise first
  const expertise = await loadUserExpertise(userId);

  // Generate UI based on expertise
  return {
    recommendations: await getRecommendations(expertise),
    layout: selectLayout(expertise.behavior_patterns),
    promotions: filterPromotions(expertise.preferences),
    navigation: prioritizeCategories(expertise.inferred_interests)
  };
}

Latency Considerations

Product experts face latency challenges that codebase experts don't:

The Problem

User Request → Load Expertise → Generate UI → Response
                    ↓
              Agent thinking time (seconds)
                    ↓
              User waiting... (bad UX)

Solutions

1. Pre-computation

// Update expertise async, not on-demand
// Pre-generate UI components during low traffic

2. Progressive Loading

// Show generic UI immediately
// Load personalized elements async
// Swap in when ready

3. Expertise Caching

// Cache hot user expertise in Redis
// Invalidate on significant changes only

4. Tiered Personalization

// Level 1: Instant (cached preferences)
// Level 2: Fast (simple inference)
// Level 3: Deep (full agent, async)

Privacy and Data Handling

Data Minimization

Only store what you need:

// Good: Store patterns, not raw data
{
  "preferred_price_range": {"min": 100, "max": 300},
  "category_affinity": {"tech": 0.8, "fashion": 0.3}
}

// Bad: Store every view with full context
{
  "views": [/* hundreds of detailed entries */]
}

User Control

Provide transparency and control:

interface UserExpertiseControls {
  viewExpertise(): UserExpertise;
  clearExpertise(): void;
  disablePersonalization(): void;
  exportData(): DataExport;
}

Retention Policies

// Decay old data
function decayOldInteractions(expertise: UserExpertise): void {
  const cutoff = daysAgo(90);
  expertise.interaction_history =
    expertise.interaction_history.filter(i => i.timestamp > cutoff);
}

When to Use Product Experts

Use Case	Good Fit?	Notes
E-commerce recommendations	Yes	High value, clear signals
Content personalization	Yes	Engagement improves
Search ranking	Yes	User-specific relevance
Simple preferences	No	Traditional settings work
Compliance-heavy domains	Maybe	Privacy concerns
Low-traffic products	No	Not enough data

Implementation Checklist

Database Setup

• User expertise JSONB column
• Action tracking table
• Index on user_id for expertise lookup

Backend Services

• Action tracking endpoint
• Expertise update worker (async)
• Expertise query API
• Cache layer (Redis)

Agent Integration

• Expertise loading prompt
• UI generation prompt
• Recommendation prompt

Frontend

• Progressive loading UI
• Skeleton states while personalizing
• Fallback to generic experience

Privacy

• Data retention policy
• User control dashboard
• Export/delete functionality

Anti-Patterns

Anti-Pattern	Problem	Solution
Sync updates	Blocks user	Async queue
Unbounded history	DB bloat	Rolling window
No fallback	Broken for new users	Default experience
Over-personalization	Filter bubble	Inject diversity
No decay	Stale preferences	Time-weighted data

Example: E-commerce Product Expert

## User Expertise Structure

preferences:
  price_sensitivity: high|medium|low
  brand_loyalty: [list of preferred brands]
  category_interests: {category: affinity_score}

behavior:
  browse_vs_buy_ratio: 0.15
  cart_abandonment_rate: 0.4
  avg_time_to_purchase: 3 days

purchase_history:
  total_orders: 12
  avg_order_value: 150
  last_purchase: 2025-01-10

## Personalization Actions

1. Show price-sensitive users sale items first
2. Highlight preferred brands in search results
3. Remind high-abandonment users of cart items
4. Suggest reorder for repeat purchases

Related Skills

• agent-expert-creation: Core expert patterns
• expertise-file-design: Schema design principles
• self-improve-prompt-design: Maintaining accuracy

---

Last Updated: 2025-12-15

Version History

• v1.0.0 (2025-12-26): Initial release

---