A trilogy. Speculative fiction — cosmic horror, political tragedy, philosophical SF. Nine human POV characters and one non-human narrator. Three books. 550,000+ words as of this writing. The prose was generated by AI. The entire creation history is in this repository.
Two model families. Two roles.
Planning: Claude Sonnet 4.5 and 4.6 handled chapter-level structure — outlines, character briefs, continuity checks, per-chapter metadata. Claude Opus 4.6 handled complex planning — worldbuilding, trilogy-level architecture, thematic design, the multi-POV structure across three books.
Prose: Claude Opus 4.6. Exclusively. Every word of generated text — dialogue, narration, description, epigraphs, in-world documents — came from Opus. No exceptions.
The distinction matters. Planning models are interchangeable at a given quality threshold. The prose model is not. Opus generates the voice. Sonnet scaffolds the structure the voice moves through.
The project uses the BMAD Method, an open-source framework for AI-assisted development. BMAD structures work through specialized agents, defined workflows, and documented artifacts. Every stage — concept, worldbuilding, character development, chapter planning, generation, revision — produces traceable output. The framework was chosen because it makes the process auditable. The audit is the repository.
OpenCode + VSCode + Windsurf IDEs with Copilot and Cascade as the orchestration layer. The agents, workflows, and generation pipelines run through this environment. It is where the conducting happens.
I did not write this book. I conducted it.
The distinction: I conceived the story, designed the characters, built the world, architected the structure across three books, planned each chapter, queued the generation, reviewed every line of output, revised what didn't work, and approved what did. The ideas are mine. The characters are mine. The thematic argument is mine. The voice is calibrated to match mine. But the prose is the system's output — directed, constrained, reviewed, and selected by me.
Specifically:
- Concept and premise — the world, the thematic core (what systems do to people), the genre positioning
- Narrative architecture — multi-POV structure, character-to-theme assignments, pacing across three books, structural decisions (the non-human narrator, the mid-trilogy pivot, the progressive depletion of Book 3's finale)
- Worldbuilding — the vertical city, the energy economy, the institutional structures, the theology, the military doctrine, the academic framework, the resistance movement
- Character design — nine POV characters plus AEGIS, each with backstory, arc, relationships, and the specific physical details that ground them
- Editorial control — selecting, revising, or regenerating output; maintaining continuity across 550,000+ words; directing tone, pacing, and emotional register chapter by chapter
- Quality assurance — internal consistency, thematic coherence, narrative logic, style audits per chapter
The system generated prose. Given a chapter plan, character dossiers, worldbuilding references, continuity notes, a style profile, and a meta-narrative framework, Claude Opus 4.6 produced the text. The text was then reviewed, accepted, revised, or regenerated. The system also performed continuity checks, style audits, and thematic tracking through the BMAD agent pipeline.
The system did not conceive the story. The system did not decide what happens. The system did not choose which characters carry which themes. The system generated language within the constraints I built and maintained.
The public repository at https://github.com/hiKareeem/untitled contains the full process:
- Every commit from inception to current state
- Agent definitions and workflow configurations
- Worldbuilding documents, character dossiers, chapter plans
- Per-chapter metadata, style audits, thematic tracking
- The complete version history — any passage can be traced back through its development
The repository is the disclosure. It is not a summary of the process. It is the process.
The repo is public. The process is documented. This file exists because I think that's how it should work.
AI-assisted creative work raises real questions — about attribution, about the nature of authorship, about what it means to make something when the language comes from a model you don't fully control. I don't have answers to all of them. I have a repository with a complete audit trail and a willingness to discuss the rest.
Under current US case law, purely AI-generated text may not be copyrightable. The selection, arrangement, and creative direction of that text by a human may be. This repository documents sustained human creative contribution at every level — from concept through architecture through editorial selection. I think that meets the threshold. The legal landscape is evolving. The record is here for whoever needs to evaluate it.
Kareem
Discord: hikareem
hi@kareemalbaba.com