diff --git a/backend/internal/model/captain_models.go b/backend/internal/model/captain_models.go
index a521aafb..e82b8b24 100644
--- a/backend/internal/model/captain_models.go
+++ b/backend/internal/model/captain_models.go
@@ -2,6 +2,7 @@ package model
 
 import (
 	"encoding/json"
+	"fmt"
 
 	"github.com/pgvector/pgvector-go"
 )
@@ -118,7 +119,10 @@ func (a *CaptainAssistant) GetConfig() (*AssistantConfig, error) {
 		return &cfg, nil
 	}
 	if err := json.Unmarshal(a.Config, &cfg); err != nil {
-		return nil, err
+		// Return an empty config rather than nil so callers that ignore the
+		// error do not dereference a nil pointer. The raw JSON may contain
+		// type mismatches (e.g. temperature as string) from older data.
+		return &cfg, fmt.Errorf("parse assistant config: %w", err)
 	}
 	var fields map[string]json.RawMessage
 	if err := json.Unmarshal(a.Config, &fields); err == nil {
diff --git a/backend/internal/service/captain_assistant_service.go b/backend/internal/service/captain_assistant_service.go
index 2d6843ab..84848da7 100644
--- a/backend/internal/service/captain_assistant_service.go
+++ b/backend/internal/service/captain_assistant_service.go
@@ -3,6 +3,7 @@ package service
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"math"
 	"strconv"
@@ -26,6 +27,7 @@ type CaptainAssistantService struct {
 	responseRepo  *repository.CaptainAssistantResponseRepo
 	llmProvider   llm.Provider
 	cache         *redis.Client
+	promptBuilder *SystemPromptBuilder
 }
 
 // NewCaptainAssistantService creates a new CaptainAssistantService.
@@ -43,6 +45,7 @@ func NewCaptainAssistantService(
 		documentRepo:  documentRepo,
 		responseRepo:  responseRepo,
 		llmProvider:   llmProvider,
+		promptBuilder: NewSystemPromptBuilder(),
 	}
 	if len(cache) > 0 {
 		svc.cache = cache[0]
@@ -646,7 +649,7 @@ func (s *CaptainAssistantService) GenerateResponse(ctx context.Context, assistan
 	// Build system prompt from assistant config and response guidelines
 	cfg, _ := assistant.GetConfig()
 	ctx = withAssistantGenerationConfig(ctx, cfg)
-	systemPrompt := buildSystemPrompt(assistant, cfg)
+	systemPrompt := s.promptBuilder.BuildAssistantPrompt(assistant, cfg)
 
 	// Build messages for LLM
 	messages := []llm.ChatMessage{
@@ -663,6 +666,10 @@ func (s *CaptainAssistantService) GenerateResponse(ctx context.Context, assistan
 
 	resp, err := s.llmProvider.ChatCompletion(ctx, req)
 	if err != nil {
+		if errors.Is(err, llm.ErrProviderNotConfigured) {
+			applogger.L().Warnf("GenerateResponse: LLM provider not configured, returning fallback")
+			return captainPlaygroundFallbackMessage, nil
+		}
 		applogger.L().Errorf("GenerateResponse LLM call: %v", err)
 		return "", fmt.Errorf("llm generation failed: %w", err)
 	}
@@ -716,7 +723,7 @@ func (s *CaptainAssistantService) generatePlaygroundLLMResponse(ctx context.Cont
 
 	cfg, _ := assistant.GetConfig()
 	ctx = withAssistantGenerationConfig(ctx, cfg)
-	messages := []llm.ChatMessage{{Role: "system", Content: buildSystemPrompt(assistant, cfg)}}
+	messages := []llm.ChatMessage{{Role: "system", Content: s.promptBuilder.BuildAssistantPrompt(assistant, cfg)}}
 	for _, message := range history {
 		if message.Role == "" || message.Content == "" {
 			continue
@@ -731,6 +738,10 @@ func (s *CaptainAssistantService) generatePlaygroundLLMResponse(ctx context.Cont
 		MaxTokens:   1024,
 	})
 	if err != nil {
+		if errors.Is(err, llm.ErrProviderNotConfigured) {
+			applogger.L().Warnf("GeneratePlaygroundResponse: LLM provider not configured, returning fallback")
+			return captainPlaygroundFallbackMessage, nil
+		}
 		applogger.L().Errorf("GeneratePlaygroundResponse LLM call: %v", err)
 		return "", fmt.Errorf("llm generation failed: %w", err)
 	}
@@ -802,21 +813,3 @@ func appendAdditionalPlaygroundMessage(history []PlaygroundMessage, current stri
 	}
 	return append(history, PlaygroundMessage{Role: "user", Content: current})
 }
-
-// buildSystemPrompt constructs the system prompt from assistant config and guidelines.
-func buildSystemPrompt(assistant *model.CaptainAssistant, cfg *model.AssistantConfig) string {
-	prompt := fmt.Sprintf("You are %s, an AI assistant.", assistant.Name)
-	if cfg.ProductName != "" {
-		prompt += fmt.Sprintf(" You represent the product: %s.", cfg.ProductName)
-	}
-	if cfg.Instructions != "" {
-		prompt += fmt.Sprintf("\nInstructions: %s", cfg.Instructions)
-	}
-	if len(assistant.ResponseGuidelines) > 0 && string(assistant.ResponseGuidelines) != "null" {
-		prompt += fmt.Sprintf("\nResponse Guidelines: %s", string(assistant.ResponseGuidelines))
-	}
-	if len(assistant.Guardrails) > 0 && string(assistant.Guardrails) != "null" {
-		prompt += fmt.Sprintf("\nGuardrails: %s", string(assistant.Guardrails))
-	}
-	return prompt
-}