zy-zmc/tianming-novel-ai-writer

What it solves

Traditional AI writing tools often struggle with long-form narratives, leading to "hallucinations" where characters' personalities drift, plot holes emerge, and early plot points are forgotten as the context window fills up. Tianming solves this by treating a novel as a managed database of states rather than a simple text stream, ensuring consistency across thousands of chapters without relying solely on the model's memory.

How it works

Instead of feeding the entire story history into the AI, the system uses a closed-loop writing process: Read State $\rightarrow$ AI Write $\rightarrow$ Gatekeeper Validation $\rightarrow$ Commit to Text $\rightarrow$ Update State $\rightarrow$ Next Chapter.

Key mechanisms include:

• 15-Dimensional Fact Snapshot: Tracks specific variables (character levels, locations, item ownership, secret knowledge) as structured data. Each chapter must output a CHANGES JSON declaration to update these values.
• Generation Gatekeepers: Six validation layers that block a chapter from being saved if it contains contradictions (e.g., a character who is afraid of fire entering a volcano) or misses required plot elements.
• Long-Distance Recall: Uses a combination of historical milestones, vector-based semantic retrieval (via local ONNX models), and state tracking to bring relevant past events back into the current prompt.
• Human-like Refinement: Employs a local chinese-roberta-tiny model to replace AI-typical phrasing with more natural language while maintaining semantic similarity.

Who it’s for

Web novel authors and creative writers who need to maintain strict logical consistency and complex world-building over extremely long narratives.

Highlights

• State-Based Writing: Moves beyond context windows by using structured data to track story state.
• Causal Validation: A six-stage gatekeeper system that prevents logical errors from entering the story.
• Automated World-Building: Tools for "book dismantling" (analyzing existing novels) and generating structured rules for world-building.
• Local AI Integration: Built-in ONNX models for embedding and text refinement to reduce API costs and improve privacy.
• .NET 8/WPF Architecture: A professional desktop application with a full-featured Markdown editor and multi-model support.