AI Characters & Conversational NPCs¶
English edition. The Japanese edition (../../ai-characters/) is the source of truth; this translation may lag behind it.
This is the first area where this knowledge base aims to be the best in the world. Scope: characters that talk through an LLM, remember, and build a relationship with the player — game NPCs, AI companions, AI VTubers, and official IP characters brought to life with AI.
Why this area¶
- It changes the most. It is the clearest example of an experience that cannot exist without AI.
- Knowledge is fragmented. Game AI research, companion apps, streaming culture and regulation are discussed in separate silos.
- Japan has an edge. Character culture, VTubers, voice actors — and, since August 2026, a Ministry of Justice guideline on voice rights.
- Regulation now shapes design directly. The scope of a character's conversation alone can change its legal classification (see pattern P01).
Contents¶
| File | What it is | Use it for |
|---|---|---|
| patterns.md | 38 design patterns in 7 groups | A reference while planning and designing |
| case-studies.md | 14 case studies (3 from Japan), successes and failures | Evidence of what has actually happened |
| evaluation.md | A five-layer evaluation framework with release gates | Development, QA and ship decisions |
| playbook-persuasion-npc.md | Playbook: building NPCs that make persuasion a game (from our own experiments) | When implementing |
Five principles (the catalog in one screen)¶
- Decide the conversation scope first. Experience, cost and legal classification all follow from it (P01).
- The engine owns facts and rules. Never let what the LLM says become game truth (P16).
- Don't let the character judge itself. Persuasion outcomes are scored by a separate judge (P17).
- Keep relationships as numeric state. Affinity must not depend on the LLM's mood (P11).
- Design the ending. Endless conversation is a weakness for wellbeing and for compliance (P22).
Experiments¶
A small harness that compares a naive NPC engine with one that applies these patterns lives in prototype/; results are logged in experiments/.