Your creativity. Your agent. Your book.
Give your AI agent an idea, even a one-line one, and Book Genesis takes it to a complete book: planned, written chapter by chapter in one voice, read blind by simulated readers, audited like a skeptical editor would, revised, and packaged with a logline, blurb, synopsis, query letter and cover brief. You agree at each step, or tell it to run to the end.
Read a book it made: Vicente, O Lago das Mensagens · 32 chapters · 51,945 words · EPUB and PDF · Brazilian Portuguese. Made with an earlier version; its score is self-assessed and no outside reader has reviewed it yet. Judge it yourself.
Version 6.0.0-beta.1 · MIT licensed · free · runs on your own agent and account.
A book written with the measured patterns of books readers finish, criticized by readers who cannot see the plan, and packaged to sell. That is the goal every part of the workflow is built around. It is not a promise: no workflow can guarantee a bestseller, and this one says so in every report.
- Intake. Your idea becomes a brief: genre, reader, length, comparable books in your idea's language, and four simulated readers built for this book.
- Foundation and architecture. Characters with wounds and contradictions, a theme asked as a question, a voice with a rhythm anchored to real books, a chapter plan, and a ledger that keeps facts straight.
- Drafting. Every chapter saved as a file. Chapter 1 goes to the four readers before the rest is written; every fifth chapter gets a continuity check and another read.
- Audit and revision. A read-only auditor quotes exact passages. Revision runs against two gates, the readers' and a ten-part rubric, at most three passes each. A revised chapter is kept only if the blind readers prefer it.
- Score and package. A plain report of where the book stands, plus the editorial package, a hook test of alternative titles and blurbs, and EPUB and PDF when your machine can make them.
Everything lives in a folder you own. Close the session, come back tomorrow in any supported agent, and it picks up from the files.
You need Python 3.10 or newer, Git, and an agent that can read and write files.
git clone https://github.com/felipelobomotta-blip/book-genesis-studio.git
cd book-genesis-studio
python runner/installer.py install claudeReplace claude with your agent. python runner/installer.py targets lists them all; --dry-run shows what would change first; verify-install checks the result. Upgrading from 5.x retires the old skills automatically when they are unchanged. Details: installation guide.
| Agent | Target | Status |
|---|---|---|
| Claude Code | claude |
first class; blind readers run as read-only subagents |
| Codex | codex |
first class; subagent isolation checked on the first run |
| OpenCode | opencode |
first class; subagent isolation checked on the first run |
| Hermes Agent | hermes |
first class; isolated subagents |
| OpenClaw | openclaw |
first class; isolated subagents on request |
| Kimi Code, Cursor, GitHub Copilot, Antigravity (app and CLI), Gemini CLI, DeepSeek Harness, Qwen Code, Pi, Windsurf, a shared folder | see targets |
installs; not yet tested with a full book |
What has actually been run where is tracked in compatibility. A full 6.0 book in each first-class agent is part of the release plan, not a claim yet.
Open your agent in an empty folder for the book and say:
Use book-genesis. My idea: a retired train dispatcher finds a farewell letter inside a station clock.
In agents with skill commands, /book-genesis works too. The agent writes in the language you use. To let it finish without stopping, say "go to the end" at any check-in.
- book-genesis: the full workflow, with its phase prompts, the three critic roles, the rubric, the patterns library and the specialists for research, character, prose, continuity, revision, production and series.
- beta-reader: three very different test readers for any draft.
- editorial-package: logline, blurb, synopsis, query letter and cover brief for any finished manuscript.
- literary-agent-panel: simulated agents, an acquiring editor, a bookseller and target readers judge market fit.
The three smaller skills work on their own, on any manuscript.
- The writer never sees the rubric or the target; the readers never see the plan or earlier scores.
- Critics run in isolated subagents, or in a second AI tool from another company when you have one installed, and every verdict prints how independent it was.
- Four readers with different stakes decide, not one: the genre fan, the reader who did not want to like it, a reader from the next shelf over, and one who gives any book ten pages at an airport.
- Revision stops after three passes per gate and says exactly what is still missing, instead of polishing until a critic agrees.
It is simulated reading, and the reports say so. Architecture: docs/architecture.md. Questions: FAQ.
Eleven projects across genres and languages, with what each one taught the system. Every case so far is the maintainer's, which is the biggest gap in this repository. Wrote a book with it? Add it through issue #13; an honest note about what worked and what did not is worth more than a star.
python runner/installer.py verify-suite
python -m unittest discover -s tests -vThe suite check refuses any skill that points outside itself, names a retired skill, or breaks the pipeline order. Edit role files under skills/book-genesis/references/roles/, then run python runner/installer.py generate-agents. See contributing, roadmap, security and the changelog.
Created by Felipe Lobo. MIT licensed.