This repo contains only the process design. The code, prompts and style rules are private.
Used in production on a full novel. The publishing project has since closed, so the tool is no longer developed.
A novelist and I tried to start a publishing house. His job was to design the story structure and the prose style for our target readers.
Once that was done, a single book of about 50 chapters (a length chosen for our target) took 3–4 months from zero to a corrected first draft.
I mapped the process, and one step took most of the time: turning the structured scene outlines into a first draft. It was also the least valuable step: going from first draft to corrected draft was much faster and fixed most of the problems, and beta readers then polished the corrections.
We tried the existing tools first:
| GPT-4o via API, Novelcrafter | What we needed | |
|---|---|---|
| Context | loses track across 50 chapters | every chapter knows the characters, places and facts already used |
| Style | follows the rules loosely | follows a strict, house-specific rulebook |
| Control | one prompt, one answer | each rule checked, and the draft sent back when it fails |
- Automated the first draft. The writer starts from a generated draft and spends his time correcting it, instead of writing everything from a blank page.
- Turned the novelist's method into precise rules. Together with him, I wrote his style and structure down as explicit rules, so that a model can follow them and a check can verify them. The rulebook kept growing as we found new problems. About 30% of the rules are checked by plain code (word counts, banned patterns); the rest are checked by a model.
- Kept the model out of the control flow. The order of the steps, the approval logic and the retry limits are fixed in code. The model only writes and reviews, plus one choice per chapter: which reviewers to run and how many rewrite rounds to allow.
flowchart TD
START(["▶ START · new chapter"]) --> IN
IN[/"📄 INPUTS<br/>Story structure and chapter summaries · Character, place and object sheets<br/>Style rulebook · Facts already used in earlier chapters"/]
IN --> S["⚙️ CODE<br/>Select the context for this chapter"]
S --> P["🤖 AGENT<br/>Plan the chapter: which reviewers, how many rounds"]
P --> D["🤖 AGENT<br/>Write the first draft"]
D --> R
subgraph R["🤖 AGENTS · review in parallel"]
direction LR
R1[Style]
R2[Continuity]
R3[Characters and<br/>point of view]
R4[Logic and<br/>reader's view]
end
R -->|changes needed| W["🤖 AGENT<br/>Rewrite only the flagged parts"]
W -->|limited rounds| R
R -->|all approve| C["⚙️ CODE<br/>Automatic rule checks"]
C --> F["🤖 AGENT<br/>Final polish, then re-checked"]
F --> L["⚙️ CODE<br/>Record the facts used in this chapter"]
L -.->|next chapter| IN
L --> H["👤 WRITER<br/>Corrects the chapter"]
H --> END(["✅ END · chapter done"])
classDef input fill:#3a3a3a,stroke:#bfbfbf,color:#fff
classDef terminal fill:#111,stroke:#fff,stroke-width:3px,color:#fff
class START,END terminal
classDef code fill:#1f4e79,stroke:#9cc3e6,color:#fff
classDef llm fill:#a3541c,stroke:#f4b183,color:#fff
classDef human fill:#2e6b30,stroke:#a9d18e,color:#fff
class IN input
class S,C,L code
class P,D,W,F,R1,R2,R3,R4 llm
class H human
How the loop stays reliable:
- Each reviewer checks one thing and never reads the other reviewers' reports, so one opinion doesn't sway the others. A chapter passes only if every reviewer approves and the automatic checks pass too.
- A missing or malformed review counts as "needs changes". A review that quotes text that isn't in the draft is thrown out. When the rewrite rounds run out, the chapter goes to the writer with a flag.
- After each chapter, the facts it used (places, objects, details) are recorded, so the next chapters don't repeat or contradict them.
How it stays cheap:
| Models | Haiku for narrow checks, Sonnet for writing and rewriting |
| Caching | the fixed part of each prompt is cached: from the second round on, most input tokens cost ~10% |
| Parallel reviews | reviewers run at the same time, since they don't depend on each other |
| Tracking | tokens and cost logged for every chapter |
| Executors | same pipeline through the API, through an agent harness with no API key, or by hand to debug a step |
| Before | After | |
|---|---|---|
| Time per book (~50 chapters) | 3–4 months | ~2 weeks: 1 week of structuring, 1 week of corrections |
| Writer's correction time, whole book | – | ~20 hours |
| What the writer spends time on | writing drafts | structuring and correcting |
Measured on one novel. The drafts still need the writer, but fewer corrections as models improve, because how well the rules are followed depends on the model.
| Tests | 175 automated tests on the pipeline |
| Evals | 15 cases built on real material, covering 11 reviewer roles. Each case lists what a good reviewer must find; every prompt change is scored against them |
| Decisions | 12 written architecture decisions, including rejected options. Example: merging all reviewers into one call was cheaper, but the reviewers influenced each other, so I dropped it |
The prompts, the rulebook and the orchestration are the core of the work and belong to the publishing project, so they stay private.