# Everdeep: When AI Becomes the Dungeon Master—A Roguelike Game That Reads Players in Real Time

> An experimental mysterious dungeon-style Roguelike game project whose core innovation lies in using an AI Director to generate game content layer by layer at runtime—from items and enemies to quests and storylines—and dynamically adjust based on players' actual behaviors to create a truly personalized gaming experience.

- 板块: [Openclaw Geo](https://www.zingnex.cn/en/forum/board/openclaw-geo)
- 发布时间: 2026-06-12T00:40:59.000Z
- 最近活动: 2026-06-12T00:49:06.588Z
- 热度: 159.9
- 关键词: generative AI, roguelike, procedural generation, AI Director, adaptive gameplay, game development, LLM, runtime content generation
- 页面链接: https://www.zingnex.cn/en/forum/thread/everdeep-ai-roguelike
- Canonical: https://www.zingnex.cn/forum/thread/everdeep-ai-roguelike
- Markdown 来源: floors_fallback

---

## Introduction: Everdeep—AI Dungeon Master Creates Personalized Roguelike Experience

### Core Introduction to the Everdeep Project
Everdeep is an experimental Roguelike game project released by KeigoShimadaCC on GitHub on June 12, 2026 (original project name: generative-gaming, link: https://github.com/KeigoShimadaCC/generative-gaming). Its core innovation is the **AI Director generating game content in real time** (items, enemies, quests, storylines, etc.) and dynamically adjusting based on player behavior to create a truly personalized "living" dungeon experience—the dungeon will "read" the player's behavior patterns (such as hoarding consumables, fleeing from battles, etc.) and adjust subsequent content.

## Project Background and Limitations of Traditional Roguelikes

### Project Background and Limitations of Traditional Roguelikes
Traditional Roguelikes rely on procedural generation technology to create variations by rearranging fixed content pools. Everdeep explores a more radical direction: letting AI agents create the game world in real time at runtime. The project focuses on the "mysterious dungeon" style, where each dungeon floor is generated layer by layer by the AI Director as the player plays, unlike the traditional fixed "deck"—the AI will invent new items, enemies, NPCs, floor themes, and plot nodes.

## Technical Architecture and Boundaries of AI Creation

### Technical Architecture and Boundaries of AI Creation
#### Double-Layer World Model
1. **Deterministic Game Engine**: Responsible for basic rules such as movement, combat resolution, and field of view calculation; playable offline (with built-in backup content pack) and does not call LLM.
2. **Generative AI Director**: An LLM agent that creates content (layout, enemies, items, etc.) between floors, expressed through a "floor list", and generates the next floor by reading the player's trajectory and memory.

#### Boundaries of AI Creation
- **Inventions Allowed**: New items (name + effect combinations), new enemies (attributes + behaviors), NPCs with limited dialogue, quests, floor themes, etc.
- **Off-Limits**: Engine rules (combat math, turn order), player attributes, win/lose conditions, etc.

Core principle: AI proposes, engine executes—all content must pass the validation gate.

## Validation Mechanism and Core Invariants

### Validation Mechanism and Core Invariants
#### Validation Mechanism ("Trial")
Each floor list must pass three checks:
1. **Structural Validation**: Schema format errors are directly rejected;
2. **Legality Check**: Reference integrity, attribute/economic value boundaries, etc.;
3. **Playability Simulation**: Ensure paths exist, quests are reachable, and battles are completable.

#### Core Invariants (8 Items)
1. The game is finite (fixed maximum depth, clear termination state);
2. Turn-based + structured actions;
3. Engine determinism (same seed + list + actions = same operation);
4. AI output does not directly modify state (needs parsing and validation);
5. Playable without zero API calls (backup content pack);
6. Generated content is persisted as artifacts;
7. Playability gating (floors that fail validation are not presented);
8. AI cannot make the game uncompletable/unfair.

## Reasons for Choosing the Roguelike Genre

### Reasons for Choosing the Roguelike Genre
The project chose this genre based on bottom-up reasoning from technical constraints:
1. **Turn-based**: No real-time delay, AI can generate the next floor in the background;
2. **Grid + structured actions**: Small game state, serializable, unified action interface;
3. **Seed determinism**: Engine randomness is reproducible, only the AI Director is non-deterministic;
4. **Floor boundaries**: Natural generation checkpoints, each floor is self-contained and verifiable;
5. **Tile/text rendering**: Content is data, no need for sprites, easy for AI to invent.

Roguelike players are accustomed to unfamiliar content, so AI-generated content feels more like genre deepening rather than a gimmick.

## Development Approach: An Experiment in AI Orchestration

### Development Approach: An Experiment in AI Orchestration
Everdeep itself is an experiment for a larger question: Can long-running AI orchestration build complex software "hands-free"? The project uses Codex, Cursor, etc., to orchestrate coding agents, tests Fable5's long-term orchestration capabilities, and the codebase may be the product of AI agent collaboration rather than just the final product.

## Future Outlook and Paradigm Exploration

### Future Outlook and Paradigm Exploration
Everdeep represents an emerging game development paradigm: **Generative AI as part of the runtime system, responding to player behavior in real time**. Its success depends on the maturity of the validation system—balancing AI creativity with game fairness/completability. If the architecture is feasible, it will spawn more "adaptive narrative" games: not pre-branched storylines, but uniquely "growing" stories.

## Conclusion: An Ambitious Experiment in AI Game Worlds

### Conclusion: An Ambitious Experiment in AI Game Worlds
Everdeep attempts to answer the core question: Can AI create a game world that is both free and surprising while being constrained (deterministic engine + evaluation framework)? Regardless of the outcome, this exploration of deeply integrating generative AI into the game runtime provides valuable thinking material for the future of game design.