重庆大学AI训练营+202524131034T+nexent-worldbuilding-agent

[1ASunny]

Project Introduction

I built Nexent Worldbuilding Agent, a structured creative workflow for designing, diagnosing, and maintaining fictional worlds in Nexent v2.4.0.

Instead of generating a large block of disconnected lore in one pass, the Agent guides creators through focused decisions, checks each decision against explicit criteria, and stores confirmed results in a versioned Markdown world bible.

Repository:
https://github.com/1ASunny/nexent-worldbuilding-agent

The Problem

Many worldbuilding tools optimize for generating more ideas: factions, species, magic rules, historical events, and place names. However, a setting may still feel shallow when those ideas are disconnected, stop after one layer of consequences, or never affect a character, scene, conflict, or player choice.

This project follows a different principle:

A small number of strong deviations should be developed deeply and connected to the actual story or gameplay.

The Agent therefore:

• keeps no more than three core deviations;
• develops each deviation through a consequence chain;
• checks compatibility and extrapolation;
• audits whether every important lore element has a narrative or interactive interface.

Why Nexent

Nexent provides the conversation, model reasoning, Progressive Skill loading, visible tool traces, and MCP integration required for a multi-turn creative workflow.

Responsibilities are deliberately separated:

• Nexent Agent: understands the creator’s goal, proposes concrete choices, explains trade-offs, and waits for confirmation.
• Progressive Skill: defines the workflow modes, thirteen-step method, and hard criteria.
• MCP workspace: saves and restores confirmed Markdown checkpoints, increments versions, and calculates SHA-256 hashes.
• Creator: remains responsible for every consequential creative decision.

flowchart LR
  U["Creator request"] --> N["Nexent Agent"]
  N --> S["read_skill_md(worldbuilding)"]
  S --> R["Diagnosis / lightweight / full workflow"]
  R --> Q["Present current choices"]
  Q --> C["Creator confirms one decision"]
  C --> M["save_worldbuilding_checkpoint"]
  M --> B["Versioned world-bible.md"]
  B --> Q

Loading

Three Workflow Modes

1. Diagnosis Mode

For an existing setting that feels thin, confusing, tonally inconsistent, disconnected from the story, or difficult to scale.

The Agent identifies the closest symptom, explains its likely cause, and points to the workflow step that should be revised.

2. Lightweight Mode

For one core deviation, one magic or progression system, one blocked design problem, or a short-form setting.

Only the relevant parts of the method are loaded instead of forcing the creator through all thirteen steps.

3. Full Pipeline

For novels, serialized fiction, animation, games, screenplays, and collaborative lore bibles that require long-term consistency.

The Agent advances one step and one creator decision at a time.

The Thirteen-Step Method

The complete workflow covers:

1. Choosing one primary entry point.
2. Defining no more than three one-sentence core deviations.
3. Selecting the intended audience experience before deciding the required worldbuilding depth.
4. Building consequence chains with a target depth of seven layers.
5. Checking compatibility between core deviations.
6. Auditing whether every consequence follows from the previous layer.
7. Allocating a mystery budget between transparent mechanisms and unexplained origins.
8. Calibrating tone across concrete cultural elements.
9. Checking narrative and gameplay-rule consistency for interactive media.
10. Auditing narrative interfaces, constraints, resources, pressures, and player choices.
11. Placing familiar anchors so readers can enter an unfamiliar world.
12. Controlling unnecessary precision and exposition density.
13. Selecting a maintenance strategy for long-running or collaborative projects.

The consequence chain is the central workload. If a deviation cannot produce a second layer, the Agent recommends replacing it. Five layers begin to produce character motivation; seven layers begin to produce a civilization’s values.

Interaction Design

The Agent never dumps all thirteen steps into one response.

Each round contains at most three questions, with two to four short and mutually exclusive options. When a creative decision is required, the Agent proposes concrete one-sentence candidates instead of vague directions.

After presenting the choices, the Agent stops and waits. It never silently selects an option or advances before the creator confirms the current decision.

Versioned Markdown Workspace

The MCP service provides two tools:

save_worldbuilding_checkpoint
load_worldbuilding_checkpoint

save_worldbuilding_checkpoint:

• validates the project name, current step, status, and Markdown size;
• writes each version atomically;
• calculates a SHA-256 hash;
• returns the project name, version, current step, hash, and optional download URL.

load_worldbuilding_checkpoint restores the latest version using the exact project name, allowing a new conversation to continue without repeating confirmed decisions.

The MCP service does not score creative quality or make worldbuilding decisions. It only provides deterministic persistence, versioning, hashing, and retrieval.

Example Usage

Start a complete project:

Help me build an expandable setting for a long-running science-fiction novel.
The project is called Glass Tide.

Diagnose an existing system:

My magic system becomes harder to control in later chapters. Diagnose the problem.

Resume a saved project:

Continue the Glass Tide project from the latest saved checkpoint.

A design-target example is included here:

[Glass Tide world bible](https://github.com/1ASunny/nexent-worldbuilding-agent/blob/main/artifacts/demo/glass-tide-world-bible.md)

Repository Structure

skill/worldbuilding/               Progressive Skill and references
tools/worldbuilding_workspace.py   versioned Markdown storage
tools/mcp_server.py                FastMCP adapter
nexent/agent-config.md             proposed Agent configuration
docs/NEXENT_SETUP.md               integration instructions
docs/MANUAL_DEMO_RUNBOOK.md        manual acceptance procedure
docs/TEST_PLAN.md                  automated and E2E gates
artifacts/demo/                    design-target example
artifacts/validation/              local validation record

Local Validation

The following checks have been completed locally:

• 8/8 automated tests passed on Python 3.12.
• The Progressive Skill archive was reproducibly packaged.
• Skill ZIP SHA-256:

70EF1F1491B6569E1B2692C46A89E8457753072A141AA58A7D72B6A3658354A3

• Skill structure and referenced resources were validated.
• Project-name path safety was tested.
• Atomic writes and version increments were tested.
• SHA-256 integrity and checkpoint restoration were tested.
• FastMCP 2.14.7 loaded both workspace tools successfully.

Reproduction commands:

python -m unittest discover -s tests -v
python scripts/package_skill.py
python scripts/package_skill.py --validate-only

Current Limitations

• Nexent v2.4.0 end-to-end conversation;
• visible read_skill_md → save/load checkpoint tool trace;
• independent client access to generated files;
• multi-tenant authorization and concurrent-writer testing;
• official Nexent Agent export and clean-tenant import.

The demo static-file service is unauthenticated and must not be used for production manuscripts. A production deployment requires tenant isolation, authenticated storage, retention and deletion policies, concurrency control, backups, and observability.

What I Learned

The central lesson was that structured creative assistance should control the process without controlling the creator.

Hard constraints are useful for preventing common failure modes, but they should remain explainable heuristics rather than opaque literary-quality scores.

Nexent’s separation of Agent reasoning, Progressive Skills, and MCP tools fits this pattern well: the Skill provides a stable method, the Agent manages collaboration, and the tool layer handles only state that must be deterministic and recoverable.

Next Steps

• Capture diagnosis, lightweight, full, save, and restore traces.
• Validate downloads from an independent user client.
• Add authenticated, tenant-isolated object storage.
• Test concurrent saves and version conflicts.
• Collect creator feedback on pacing, choice quality, and world-bible usefulness.

I would appreciate feedback from the Nexent community, especially on Progressive Skill loading, multi-turn interaction design, checkpoint recovery, and production storage patterns.

License and Attribution

The project is released under the MIT License.

The bundled Worldbuilding Skill is adapted from [tahini2007/worldbuilding-skill](https://github.com/tahini2007/worldbuilding-skill), with its license and attribution preserved.