Candidates: try freely, discard freely
Dry runs happen on a copy of the project. A candidate that fails a check goes back for revision; you can keep adjusting or keep the original.
How it works / VERSION 6.0
How material flows, how reviews shape what comes next, and how every change is approved and saved. This page walks through each part of COSMO.
01 / Components
The author sets the direction and makes the calls. The AI researches, drafts and revises. Local tools connect preparation, checking and saving into one workflow, and whatever is adopted is written into the book's own files.
Co-writing
Workflow
The book's record
Every formal operation runs through novel.py. The model and the workflow are independent: another AI tool that works with local files can continue from the book's entry files and current state.
02 / Chapter loop
Setting up a book starts with the material you already have: story direction, structure and voice are worked out and confirmed before initialization. Writing then runs in batches of three chapters by default, one at a time if you prefer, up to five. Every batch follows the same loop.
Set the focus and the changes from the current outline.
Bring in setting, voice, character state and related chapters.
Two passes, one revision, a full reread, then an independent review.
List every change and check content against records.
Chapters and records become part of the book together.
Dry runs happen on a copy of the project. A candidate that fails a check goes back for revision; you can keep adjusting or keep the original.
Only changes you approve, and that pass the final pre-save check, are written to the book. Explorations and rejected options never count toward the manuscript, progress or character facts.
Voice samples also go through a candidate first: INIT during book setup, REVISE once the book is under way, written only after the sample analysis is complete and you approve. The two passes focus on reader momentum, then continuity and language; each batch also gets an independent review, and the whole arc is reread when it ends.
03 / Stage reviews · 6.0
Reviews count chapters in finalized reading order. Chapters three, six and nine look back at recent reading experience; chapter twelve reviews the structure of the stage and includes the short review due then.
Are earlier expectations moving forward, or is the story setting things up again and again without a result?
With a new setting, object or opponent, does the reader still get the same feeling? Do quiet or repeated moments add something new?
Share a stage summary and hear your thoughts on direction and motivation; decide from the reading evidence what to keep or adjust, and check the effect next round.
Name the reading problem before choosing a fix. Never pile on conflict, twists or cliffhangers mechanically.
Read the clean manuscript in batches and note real reactions first, then compare against the main promise, chapter cards and earlier assessments, rather than explaining everything from the plan.
Research into ranked books is optional. If you do it, pick two or three works from relevant lists and prefer three legally accessible consecutive chapters or a complete scene.
Based on the reading and your feedback, write down why the plan stays or what changes. Borrow general techniques, never specific plots or phrasing.
When research is done, record list sources, dates, works and reading scope; when it is skipped, note the reason. If access is limited, record what was actually read; no follow-up task is added automatically.
Book setup uses INIT; books under way use REVISE. Agreed changes go into the original outline and related chapter cards before shaping later writing, and the scope of each chapter candidate stays unchanged.
Each review records the chapter order and manuscript digest it covered. If an earlier chapter is edited or reordered, the system asks for a fresh check. Preparing a task, refreshing material or logging a deferral never counts as a completed review.
The program identifies reviews that are due and records what they covered. Waiting for your reply does not count as a conversation, and an existing reply for this round is picked up directly. Local fixes and stage-level judgments are reported separately; the quality of reading, conversation and research still needs real review.
04 / Approve and save
The system calculates a content digest for every candidate and records the state of the related files during the dry run. Everything is checked again before saving, so a draft cannot change after you approve it.
The draft and scope of changes you reviewed
Same as when you approved it
Same as during the dry run
Once the digest matches and the required checks pass, the chapters and related records are written together.
If the draft becomes version B, rerun the dry run and approve the new digest. If the related book files change, recheck against the current files, then rerun the dry run and approve.
The content digest pins down exactly what you approved. It cannot prove you read every word, and it does not replace reviewing characters, language and cause and effect.
05 / Run feedback
Misunderstandings, rework and workflow blockers are written up by the AI from the visible conversation and existing evidence: what happened, how it was judged, how it was handled and how it was verified.
Record conversation excerpts, tool output or the AI's observations, with their source and visible scope.
Separate creative choices, execution slips, template limits and software bugs, and keep explanations that still need testing.
Record the correction and the steps it affects; if the system is left unchanged, record why.
Open issues carry into the next round and are checked against real revisions and reader feedback; a recurrence keeps the same number and its full history.
Run records live in the book's append-only log and are included in backups. They never alter creative commit records and are never pasted wholesale into chapter reading packages.
The system never scrapes all your chats or other tasks; the AI records only evidence it actually saw. When the system is updated, only reusable methods and rules are extracted.
06 / Isolation and updates
The downloaded folder is ready to hold one book. Keep a blank copy before you start and duplicate it for each new book, or download a fresh copy. Existing books keep their folders; before any tool update, the system checks the installation record and local changes.
Clean template
Shared rules, blank templates, checks and programs.
One project per book
This book's setting, manuscript, research, feedback and customizations.
The system keeps the unfinished state and blocks further writes. Only after you review the affected files and explicitly approve recovery does it roll back to the state before the save.
It completes the commit record and history from the point of interruption and keeps the verified content. Checking the status never triggers recovery.
When a write error can be caught, the system attempts to roll it back automatically; recovery after a forced interruption must be run explicitly, after checking the related records.
After a chapter commit succeeds, the candidate is moved whole into a retained-candidate directory and its location is reported. Anything saved during the commit can be checked there; it never becomes the manuscript automatically and is not part of the next approved candidate.
Candidates, retained candidates, reading packages and transaction history are excluded, so save any unused material you want to keep separately. Snapshots taken during an interrupted save are labeled separately, are not guaranteed to be consistent, and do not replace the original project's recovery records.
6.0 keeps schema 2 and the existing INIT / REVISE workflows. New books default to General mode, built around reader expectations and stage development; Progression and Conflict mode adds requirements for opponents and escalation. Cards from earlier versions keep their original meaning, and older books that never enabled the feature do not need to fill anything in. Upgrading never reopens an existing book or adopts a new story plan on its own.
07 / Limits
Technical safeguards make files, versions and records more reliable. Whether a book comes alive is still decided by real passages and the experience of reading them.
Passing software checks means only that the corresponding workflow checks passed. Literary quality and long-term benefits to your writing need real review, reader feedback and continued use.