diff --git a/CLAUDE.md b/CLAUDE.md
index 462157d..cbeeea7 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -102,12 +102,36 @@ LOWER(r.country_code) = ?
 
 Valid inputs: `"korea"`, `"KR"`, `"seoul"`, `"ap-northeast-2"`, `"icn"`
 
-### AI Prompt Strategy (`ai.ts`)
+### AI Prompt Strategy (`recommend.ts`)
 
 - Uses OpenAI GPT-4o-mini via Cloudflare AI Gateway (bypasses regional restrictions)
 - Server list format: `[server_id=XXXX] Provider Name...` for accurate ID extraction
 - Scoring: Cost efficiency (40%) + Capacity fit (30%) + Scalability (30%)
 - Capacity response in Korean for Korean users
+- **Prompt injection protection**: User inputs sanitized via `sanitizeForAIPrompt()`
+- **Token optimization**: Candidates limited to top 15 by price
+
+### Security Features
+
+- **Input validation**: Comprehensive checks with length limits (tech_stack ≤20, use_case ≤500 chars)
+- **Rate limiting**: 60 req/min per IP with in-memory fallback when KV unavailable
+- **SQL injection prevention**: All queries use parameterized statements
+- **Security headers**: CSP, HSTS, X-Frame-Options, X-Content-Type-Options
+- **API key protection**: Keys never logged, sanitized from error messages
+
+### Configuration (`config.ts`)
+
+Centralized limits and constants:
+```typescript
+LIMITS = {
+  MAX_REQUEST_BODY_BYTES: 10240,    // 10KB
+  CACHE_TTL_SECONDS: 300,            // 5 minutes
+  RATE_LIMIT_MAX_REQUESTS: 60,       // per minute
+  MAX_AI_CANDIDATES: 15,             // reduce API cost
+  MAX_TECH_STACK: 20,
+  MAX_USE_CASE_LENGTH: 500,
+}
+```
 
 ## Bindings (wrangler.toml)
 
@@ -154,9 +178,28 @@ curl -X POST https://server-recommend.kappa-d8e.workers.dev/api/recommend \
 
 ## Recent Changes
 
-- **Modular architecture**: Split from single 2370-line file into 7 modules
+### Architecture
+- **Modular architecture**: Split from single 2370-line file into organized modules
+- **Centralized config**: All magic numbers moved to `LIMITS` in config.ts
+- **Type safety**: `parseAIResponse(unknown)` with proper type guards
+
+### Features
 - **DB workload multiplier**: Database resource calculation based on use_case
-- **KV caching**: 5-minute cache with smart invalidation (empty results not cached)
-- **OpenAI integration**: GPT-4o-mini via AI Gateway for better recommendations
 - **Bandwidth estimation**: Automatic bandwidth category detection for provider filtering
 - **Tech specs update**: Realistic vcpu_per_users values for 150+ technologies
+
+### Security
+- **Prompt injection protection**: `sanitizeForAIPrompt()` filters malicious patterns
+- **Rate limiting fallback**: In-memory Map when KV unavailable
+- **Input sanitization**: All user inputs validated and length-limited
+
+### Performance
+- **O(1) VPS lookup**: Map-based benchmark matching (was O(n×m))
+- **AI token optimization**: Candidates limited to 15 (was 50)
+- **KV caching**: 5-minute TTL, empty results not cached
+- **Parallel queries**: Promise.all for independent DB operations
+
+### Code Quality
+- **Dead code removed**: Unused `queryVPSBenchmarks` function deleted
+- **DRY**: `DEFAULT_REGION_FILTER_SQL` shared between handlers
+- **Name-based filtering**: Provider queries use names, not hardcoded IDs