Skip to main content

Introduction to LibriScribe

LibriScribe is an AI-powered book writing assistant designed to streamline the creative process. It uses a multi-agent architecture, with each agent specialized in a particular task. It now supports three setup flows: Simple Guided Setup, Advanced Guided Setup, and Expert configuration files for repeatable runs and deeper customization.

Core Concepts​

LibriScribe is built around the idea of a multi-agent system. Each agent is a Python class responsible for a specific aspect of the book writing process. This modular design makes the project extensible and easier to maintain.

Agents​

Here's a brief overview of the key agents:

  • ProjectManagerAgent: Manages the overall workflow and coordinates the other agents. This is the main interface for the command-line tool.
  • ConceptGeneratorAgent: Generates initial book concepts, including title, logline, and a detailed description.
  • OutlinerAgent: Creates a comprehensive chapter-by-chapter outline for the book.
  • CharacterGeneratorAgent: Generates detailed character profiles, including background, personality, and relationships.
  • WorldbuildingAgent: Creates detailed worldbuilding information (history, culture, geography, etc.) relevant to the book's genre and setting.
  • ChapterWriterAgent: Writes the first draft of a chapter scene-by-scene, with narrative constraint injection and quality loop.
  • EditorAgent: Refines and edits a chapter, focusing on clarity, consistency, grammar, and style.
  • StyleEditorAgent: Refines the chapter's writing style based on specified tone and target audience preferences.
  • ContentReviewerAgent: Reviews chapter content for consistency, clarity, and plot holes.
  • FactCheckerAgent: (Primarily for non-fiction) Verifies factual claims made within a chapter.
  • PlagiarismCheckerAgent: Identifies potential plagiarism issues in a chapter.
  • ResearcherAgent: Conducts web research on a specified topic and provides a summary of findings.
  • FormattingAgent: Combines all generated chapters into a single, well-formatted Markdown or PDF document.

Narrative Quality Layer​

LibriScribe 0.5.0 includes three systems that close the loop between writing and quality automatically:

  • NarrativeGraphBuilder: Extracts structured narrative facts (entities, predicates, values) from each chapter after it is written. Facts are persisted in narrative_graph.json and survive partial runs.
  • InvariantChecker: Before each scene is generated, detects hard and soft violations (dead characters reappearing, physical actions by injured characters, destroyed locations revisited) and injects a NARRATIVE CONSTRAINTS: block into the scene prompt.
  • ContentQualityAgent: Scores prose across 5 axes — cliche density, show-don't-tell ratio, dialogue voice consistency, sentence variety, and scene function clarity. Operates in two modes:
    • Option A (Preventive): Prior-chapter quality flags injected into the next chapter's scene prompts as STYLE CONSTRAINTS:.
    • Option B (Reactive): Per-scene inline rewrite when the overall score falls below 0.65, targeting only the flagged issues while preserving all plot facts.

Getting Started​

See the Installation Guide for detailed instructions on setting up LibriScribe.

Usage​

The Usage Guide explains the current CLI workflow, including Simple, Advanced, and Expert modes, plus provider defaults, project-level model overrides, and per-agent model selection.