Lyuyuan AI Lyuyuan AI
Lyuyuan AI
Status: public historical prototype.
One-line positioning: an LLM-driven narrative RPG prototype where characters speak, remember relationship state, and trigger events that can be saved and inspected.
Why
I built this as a product-shaped agent prototype: instead of a one-off chat demo, I wanted characters, relationship state, events, saves, and a UI that could support repeated interaction.
What I Built
- Multiple selectable AI characters with distinct persona prompts.
- A two-channel response pattern: character-facing reply plus internal analysis JSON.
- Relationship, emotion, and event-state updates across turns.
- Event bus architecture for logs, statistics, achievements, and lifecycle events.
- JSON save slots, API routes, tests, and a Flask-based web UI.
Hard Parts
- Keeping character dialogue expressive while extracting structured internal state reliably.
- Separating story events, relationship updates, logging, and UI actions so the prototype could grow without turning into one large callback chain.
Evidence
- GitHub repository
- Public repo includes API/design notes, tests, version history, character configuration, and save/load behavior.
Result / Lesson
The project taught me that narrative agents need boring infrastructure: state schemas, events, logs, and save files. Without those, good dialogue is hard to debug or improve over time.
Lyuyuan AI
状态：公开历史原型。
一句话定位：一个 LLM 驱动的叙事 RPG 原型，角色可以对话、维护关系状态，并触发可保存、可检查的事件。
Why
我把它当成一个产品形态的 agent 原型来做：不只是一次性聊天 demo，而是希望有角色、关系状态、事件、存档和可以反复交互的 UI。
What I Built
- 多个可选 AI 角色，每个角色有独立 persona prompt。
- 双通道响应模式：面向玩家的角色回复，以及内部分析 JSON。
- 跨回合维护关系、情绪和事件状态。
- 事件总线架构支持日志、统计、成就和生命周期事件。
- JSON 存档槽、API routes、测试和 Flask Web UI。
Hard Parts
- 在保持角色回复自然的同时，稳定抽取结构化内部状态。
- 把剧情事件、关系更新、日志和 UI 行为拆开，避免原型膨胀成一个巨大的回调链。
Evidence
- GitHub 仓库
- 公开仓库包含 API/设计说明、测试、版本历史、角色配置和存档逻辑。
Result / Lesson
这个项目让我意识到 narrative agents 很需要“无聊但重要”的基础设施：状态 schema、事件、日志和存档。没有这些，再好的对话也很难调试或长期改进。