- Use structured outputs (JSON)
- Implement streaming
- Design for retry logic
- Plan for model switching
- Cache aggressively

// Modern AI product pattern (2025)

interface AIFeature {
  // Streaming for responsiveness
  async *stream(prompt: string): AsyncGenerator<string> {
    const cached = await checkCache(prompt);
    if (cached) return cached;
    
    // Route to appropriate model
    const model = this.selectModel(prompt);
    
    for await (const chunk of model.stream(prompt)) {
      yield chunk;
    }
  }
  
  // Model selection (cost optimization)
  selectModel(prompt: string): Model {
    if (this.isSimple(prompt)) {
      return this.smallModel; // Fast, cheap
    } else {
      return this.largeModel; // Smart, expensive
    }
  }
  
  // Retry logic (reliability)
  async withRetry<T>(fn: () => Promise<T>): Promise<T> {
    for (let i = 0; i < 3; i++) {
      try {
        return await fn();
      } catch (e) {
        if (i === 2) throw e;
        await sleep(Math.pow(2, i) * 1000);
      }
    }
  }
}

# AI Cost Analysis: [Feature]

## Current Usage
- Daily requests: [X]
- Model: [GPT-4/Claude/etc.]
- Cost per 1K requests: [$X]
- Monthly cost: [$Y]

## Optimization Plan

### 1. Caching (Est. 80% hit rate)
- Before: [100]% paid calls
- After: [20]% paid calls
- Savings: [80]%

### 2. Model Routing
- Simple queries ([60]%): Small model
- Complex queries ([40]%): Large model
- Savings: [50]%

### 3. Batching
- Real-time: [X]% of requests
- Batchable: [Y]% of requests
- Savings: [Z]%

## Projected Cost
- Before optimization: [$X/month]
- After optimization: [$Y/month]
- Reduction: [Z]%