malhajri07/add-ai-endpoint

add-ai-endpoint

Scaffolds a new AI-powered API endpoint backed by the Anthropic Claude API. Handles system prompt design, structured output parsing, token usage tracking, and per-user rate limiting.

Inputs to gather

• Feature name — e.g., "property-description-generator", "lead-scorer", "followup-writer"
• Input shape — what the endpoint receives (e.g., leadId, propertyId, free text)
• Output shape — what it returns (e.g., generated text, score object, match list)
• System prompt context — domain knowledge the AI needs (Saudi real estate, Arabic language, REGA rules)
• Model — claude-sonnet-4-6 (fast/cheap) or claude-opus-4-6 (best quality)

Steps

1. Check AI service layer exists. Read apps/api/libs/ai-service.ts. If it doesn't exist, create it:

   // Anthropic client singleton, retry logic, token tracker
   import Anthropic from "@anthropic-ai/sdk";
   const client = new Anthropic({ apiKey: process.env.ANTHROPIC_API_KEY });
   

2. Create the route file at apps/api/routes/ai-{feature}.ts:

   • Import authenticateToken middleware
   • Validate input with zod
   • Build messages array with system prompt + user content
   • Call client.messages.create() with structured output
   • Parse response, return JSON
   • Track token usage: prompt_tokens, completion_tokens in ai_usage table

3. Design the system prompt. Must include:

   • Role: "You are a Saudi real estate expert assistant for عقاركم platform"
   • Language: "Always respond in Saudi Arabic (العربية السعودية)"
   • Domain constraints: REGA compliance, FAL license awareness, SAR currency
   • Output format: specify JSON schema if structured output needed

4. Add rate limiting. In the route:

   • Check ai_usage table: user's token count today
   • If > daily limit (default 50K tokens), return 429
   • Track usage after successful call

5. Register the route in apps/api/index.ts:

   import aiFeatureRoutes from "./routes/ai-{feature}";
   app.use("/api/ai", aiFeatureRoutes);
   

6. Create react-query hook using /add-react-query for the frontend.

7. Add API key to .env:

   ANTHROPIC_API_KEY=sk-ant-...
   

Verification checklist

• [ ] Endpoint returns correct structured output
• [ ] System prompt produces Arabic responses
• [ ] Token usage is tracked per request
• [ ] Rate limiting rejects over-limit requests with 429
• [ ] Error handling: API timeout, invalid response, rate limit
• [ ] /typecheck passes

Anti-patterns

• Don't hardcode API keys — always use environment variables
• Don't stream responses for simple endpoints — use non-streaming for structured output
• Don't skip token tracking — you'll have no cost visibility
• Don't use opus for high-volume endpoints — use sonnet for speed/cost
• Don't let unauthenticated users hit AI endpoints (except chatbot)