Google Business Profile Review MCP Server

# Using MCP Resources in Prompts This guide shows different approaches for incorporating MCP resources into prompts. ## Approach 1: Instruct AI to Access Resources (Recommended) The prompt instructs the AI assistant to read the resource themselves. This is the **cleanest and most MCP-idiomatic approach**. ### Example: `manageReviews.ts` ```typescript export function createManageReviewsPrompt() { return { name: 'manage_pending_reviews', description: 'Instructs AI to use MCP resources', handler: async (args: any) => { const prompt = `You are managing Google Business Profile reviews. **YOUR TASK:** 1. First, read the "reviews://all" MCP resource to see all pending reviews 2. For each review, generate a personalized reply 3. Use the 'post_reply' tool to submit replies **AVAILABLE RESOURCES:** - reviews://all - All unreplied reviews - locations://list - Business locations - business_profile://profile - Brand guidelines - review_templates://templates - Template examples Please start by reading the reviews://all resource.`; return prompt; } }; } ``` **Benefits:** - ✅ Clean separation of concerns - ✅ AI can access resources dynamically - ✅ Resources stay up-to-date (no stale data) - ✅ AI can choose which resources to read - ✅ Follows MCP best practices **How it works:** 1. Prompt tells AI which resources are available 2. AI reads the resources using MCP protocol 3. AI uses the data to complete the task --- ## Approach 2: Embed Resource Data in Prompt The prompt handler fetches and embeds the resource data directly into the prompt text. ### Example: ```typescript export function createManageReviewsPrompt(reviewService: IReviewService) { return { handler: async (args: any) => { // Fetch data directly const result = await reviewService.getReviews(locationName); const reviews = result.data?.reviews || []; // Embed in prompt let prompt = `You have ${reviews.length} pending reviews:\n\n`; reviews.forEach((review, i) => { prompt += `Review #${i + 1}:\n`; prompt += `Rating: ${review.starRating}\n`; prompt += `Text: "${review.comment}"\n\n`; }); prompt += `Generate replies for each review.`; return prompt; } }; } ``` **Benefits:** - ✅ Simpler for AI (data already in context) - ✅ Works with AI that doesn't support resource access - ✅ All data in one place **Drawbacks:** - ❌ Prompt can become very long with lots of data - ❌ Requires prompt to handle service errors - ❌ Data is stale (snapshot at prompt generation time) - ❌ Duplicates resource functionality **When to use:** - Legacy AI systems without MCP resource support - When you need to pre-filter or transform the data - When data is small and static --- ## Approach 3: Hybrid - Reference Resources with Optional Fallback Instruct AI to use resources, but include minimal context as fallback. ### Example: ```typescript export function createReviewResponsePrompt() { return { handler: async (context: ReviewResponseContext) => { const prompt = `Generate a reply for this ${context.starRating}-star review. **REVIEW:** "${context.reviewText}" **RESOURCES FOR ADDITIONAL CONTEXT:** - business_profile://profile - Brand voice guidelines - review_templates://templates - Example responses **TIP:** Access the resources above for brand-specific guidance. Write a personalized response:`; return prompt; } }; } ``` **Benefits:** - ✅ AI can access resources if needed - ✅ Works without resources (has basic context) - ✅ Flexible and adaptive --- ## Comparison Table | Approach | Prompt Size | Real-time Data | MCP-Idiomatic | AI Complexity | |----------|-------------|----------------|---------------|---------------| | **Instruct AI** | Small | ✅ Yes | ✅ Yes | Medium | | **Embed Data** | Large | ❌ No | ❌ No | Low | | **Hybrid** | Medium | ⚠️ Partial | ✅ Yes | Low | --- ## Best Practices ### 1. **Use Clear Resource URIs** ```typescript // ✅ Good - Clear and descriptive "Read the reviews://all resource for pending reviews" // ❌ Bad - Vague "Use the reviews resource" ``` ### 2. **Explain What Each Resource Contains** ```typescript const prompt = ` **AVAILABLE RESOURCES:** - reviews://all - All unreplied reviews with ratings and text - business_profile://profile - Your brand voice, tone guidelines - review_templates://templates - Example responses by rating `; ``` ### 3. **Provide Instructions for Resource Usage** ```typescript const prompt = ` Step 1: Read reviews://all to see pending reviews Step 2: For each review, consider the rating and text Step 3: Generate personalized replies `; ``` ### 4. **Handle Cases When Resources Might Be Empty** ```typescript const prompt = ` Read reviews://all. If there are no pending reviews, inform the user. Otherwise, generate replies for each review. `; ``` ### 5. **Don't Duplicate Resource Data** ```typescript // ❌ Bad - Duplicates resource functionality const reviews = await reviewService.getReviews(); const prompt = `Here are the reviews: ${JSON.stringify(reviews)}`; // ✅ Good - References resource const prompt = `Read the reviews://all resource to see pending reviews`; ``` --- ## Real-World Example: Updated `manageReviews` Prompt **Before (Old Approach):** ```typescript // Fetched and embedded reviews directly const result = await reviewService.getReviews(location); const reviews = result.data?.reviews || []; let prompt = `You have ${reviews.length} reviews:\n`; reviews.forEach(r => { prompt += `Review: ${r.comment}\n`; }); ``` **After (Resource-Based Approach):** ```typescript // Instructs AI to read the resource const prompt = ` 1. Read the reviews://all MCP resource 2. Generate replies for each review 3. Post using the post_reply tool Available resources: - reviews://all - business_profile://profile - review_templates://templates `; ``` **Results:** - Prompt size: 500KB → 1KB (500x smaller!) - Real-time data: No → Yes - MCP compliance: No → Yes - Maintainability: Low → High --- ## Testing Your Prompts ### Test with MCP Inspector (if available) ```bash # Start your MCP server npm run start # In another terminal, test resource access mcp-inspector read-resource "reviews://all" ``` ### Test Prompts Programmatically ```typescript // Test that prompt instructs resource usage const prompt = await manageReviewsPrompt.handler({}); console.log(prompt); // Should contain resource URIs assert(prompt.includes('reviews://all')); assert(prompt.includes('business_profile://profile')); ``` --- ## Conclusion **For new prompts:** Use **Approach 1** (Instruct AI to Access Resources) - Clean, MCP-idiomatic, real-time data - Works great with modern AI assistants **For legacy compatibility:** Use **Approach 3** (Hybrid) - References resources but includes basic context - Works with both old and new systems **Avoid Approach 2** unless you have a specific reason: - Creates large prompts - Duplicates resource functionality - Stale data