Best Practices Advanced

Shipping AI NPCs in a production game requires solving challenges that go far beyond the initial prototype: managing API costs at scale, ensuring content safety, handling network failures gracefully, and testing non-deterministic systems. This lesson covers the essential production patterns.

Latency Budget

Cost Management

• Use smaller models (Claude Haiku, GPT-4o-mini) for simple NPCs; reserve larger models for key characters
• Cache common responses and greetings to avoid redundant API calls
• Implement token budgets per NPC per session
• Use local models for offline/single-player modes

Fallback Systems

Content Safety

• Implement input filtering before sending player text to the LLM
• Validate all LLM outputs against content policies before displaying
• Log flagged interactions for review (with privacy compliance)
• Provide player reporting mechanisms for inappropriate NPC behavior

Testing Non-Deterministic Systems

AI NPCs produce different responses each time, making traditional assertion-based testing insufficient. Use LLM-as-judge evaluation: run hundreds of simulated conversations and have a separate LLM score each response for persona consistency, lore accuracy, and content safety.