forsonny/claude-code-novel-writer

Purpose and entry points

This is a Claude Code–oriented autonomous novel-writing workspace, not a conventional application service. Claude is the primary runtime and executes repository instructions to plan, generate, review, and maintain a long-form manuscript. Bootstrap and validation start at launch-novel.sh and verify-system.sh; the operational contract is centered on CLAUDE.md. Human operators can observe state through automation/dashboard.py.

Orchestration and agent layer

CLAUDE.md acts as the master orchestrator: it directs workflow decisions and delegates specialized work. Agent instruction files form the functional AI pipeline:

• Story structure and adaptive pacing: .claude/agents/plot-architect.md, .claude/agents/smart-planner.md

• Prose production and continuity review: .claude/agents/chapter-writer.md, .claude/agents/continuity-editor.md

• Domain enrichment and remediation: .claude/agents/worldbuilder.md, .claude/agents/character-developer.md, .claude/agents/error-recovery.md

Claude configuration and injected reminders reside in .claude/settings.json and .claude/context-injection.txt.

Canonical story state

The system persists narrative state as repository-managed JSON rather than a database. Planning is the principal control input: planning/novel-outline.json defines intended structure, while planning/chapter-status.json, planning/plot-progress.json, and planning/scene-tracker.json record execution progress. The orchestrator and planning agents should read these before generation and update them after completed work, making them the synchronization boundary between Claude sessions.

Narrative knowledge domains

Character and world facts are separated from planning state to support consistency checks and targeted agent work. characters/character-knowledge.json is the character/arc reference; worldbuilding/world-state.json captures setting rules and world changes. Manuscript-level metadata belongs in manuscript/metadata.json. Generated chapter content is implied to live under manuscript/, although no chapter files are present in the supplied tree.

Generation and quality loop

The intended loop is: initialize/health-check, synchronize state, assess plan and quality, generate a chapter, run checks, persist progress, and repeat. Chapter formatting and prose expectations are supplied by templates/chapter-template.md and output-styles/autonomous-novelist.md. automation/chapter-completed.sh is the likely post-generation integration boundary, with automation/quality-check.sh and automation/duplicate-checker.sh providing content validation.

Automation, recovery, and observability

Shell automation implements workspace maintenance around Claude-driven generation: session lifecycle in automation/session-init.sh and automation/session-summary.sh; state synchronization in automation/auto-sync-progress.sh; health and performance checks in automation/system-health-check.sh and automation/performance-monitor.sh. Recovery and durability are handled through automation/auto-backup.sh, automation/pre-compact-backup.sh, and automation/backup.sh. These scripts operate on local files; no external database, API, or deployed infrastructure is evidenced.

Runtime and operational boundaries

Required runtimes are a Unix-like shell, Claude Code CLI, and Python 3 for automation/dashboard.py. The dashboard is an observer/reporting interface over local project artifacts, including monitor and sync-report modes. Documentation is maintained separately in Documentation/System-Architecture.md and Documentation/User-Guide.md; it describes operation but is not part of the execution path.