Expert overview
This manual assumes a reader who has written at length, has handled git in a terminal, and has handed writing or code to Claude Code or Codex before. You don’t need fiction craft or publishing procedure memorized. How the 4 Tiers mesh and how continuity is kept are unpacked in the sections that follow.
What this section covers is not a command reference but “why it’s shaped this way.” Pass through the reasoning once, then move to architecture, and genres and verification will read as something more.
The problem in one line
Section titled “The problem in one line”Hand a book to AI and one chapter comes out fine. The problem is length. As chapters pile up, the tone wavers, the character setup defined in Chapter 3 contradicts Chapter 12, the sentence rhythm flattens, and em-dashes and generic adjectives fill the density. When the chat context grows long, the AI forgets earlier chapters, and when a session breaks, where you left off disappears. book-studio is a starting point with a gate or a file handoff planted at each of those spots.
The 4-Tier pipeline
Section titled “The 4-Tier pipeline”Writing flows in four stages. Each Tier is owned by a dedicated agent and hands off to the next through an output file.
| Tier | Agent | Output | Gate |
|---|---|---|---|
| 1. Direction | direction-agent | 01_direction_brief.md (book_type, logline, persona, tone, differentiation) | user approval |
| 2. Planning | planning-agent | 02_world_building.md, 03_character_design.md, 04_plot_architecture.md, 04_chapter_outline.json (all genres, content branches by genre) | user approval |
| 3. Writing | writing-agent | chapters/ch{NN}_draft.md, chapters/ch{NN}_meta.json (one chapter at a time) | per-chapter AI-smell gate |
| 4. Quality | quality-agent | chapters/ch{NN}_ai_check.md, chapters/ch{NN}_review.md, chapters/ch{NN}_review.json | finished on passing the gate |
The point is that the source of truth is _workspace/{book_slug}/ files, not the context. Even if a session breaks, you recover your place from progress.md and the per-stage outputs. “Rewrite ch3” jumps to Tier 3 and rewrites just that chapter, and “continue from yesterday” recovers your place from progress.md.
book_type is the routing key
Section titled “book_type is the routing key”book_type is the key that decides which skills activate. The same orchestrator turns on a different skill graph depending on the book type.
fiction: world-building (the world), character-design (characters), plot-architecture, prose-writing, dialogue-writing, continuity-checkpractical: world-building (domain background), character-design (reader persona), plot-architecture, prose-writing, editing, continuity-checkessay: world-building (context), character-design (narrator and figures), plot-architecture, prose-writing, editing, continuity-checktechnical: world-building (technical background), character-design (reader prior knowledge), plot-architecture, prose-writing, editing (technical accuracy), continuity-check
Every type shares the three Tier 2 skills (world-building, character-design, plot-architecture) along with prose-writing, editing, and ai-pattern-check. The same skills run for every type; only the content branches by genre. Fiction additionally turns on dialogue-writing and illustration. The detailed split is in genres. If the type changes mid-way, it rolls back to Tier 1.
The per-chapter AI-smell gate
Section titled “The per-chapter AI-smell gate”Every chapter has to clear the 14 categories of ai-pattern-check to be marked finished. It catches spots like template phrases, list patterns, a missing opening rhythm, and a missing experience signature, and scores each chapter 0–100, where lower is better. A score of 61 or higher fails, which blocks the chapter, and writing-agent rewrites. It’s the gate proven in blog-studio, adapted to the book context, so it looks at longer paragraphs and scene continuity together.
Continuity is an index, not memory
Section titled “Continuity is an index, not memory”Character appearance, world rules, and timeline events are extracted into consistency_index.json. Each new chapter cross-checks against this index. “A character with blue eyes in Chapter 3 can’t turn brown in Chapter 12” is the promise of this spot. Because it verifies against facts planted in a file rather than leaning on the AI’s memory, the setup doesn’t drift even as the context grows long.
Many books in parallel
Section titled “Many books in parallel”You run several books at once on one channel. Each book is isolated in its own _workspace/{book_slug}/ folder, and the root books.json is the index. active_slug points at the active book, and you switch with “continue with book X.” One book’s work doesn’t bleed its context into another. It’s the spot where writing one novel and one technical book alternately doesn’t mix the setups.
It does not auto-publish
Section titled “It does not auto-publish”The pipeline only produces per-chapter outputs. Combining the chapters that passed into a single manuscript is not an automated step; the user gathers them by hand. It does not upload to a submission or ebook platform automatically either (S8). Publishing is a spot the user decides and runs, so the manual too covers only up to assembly. That’s why there’s no publish or distribution section.
The automation entry point points at one place
Section titled “The automation entry point points at one place”.claude/agents/ and .claude/skills/ are the Claude Code storage location, and AGENTS.md is the Codex entry point. Both read AI_AUTOMATION.md. The 4-Tier definitions, the book_type skill routing, the 14 categories, the directory contract, and the security baseline S1–S8 live in that one file, so the Tier and output contracts are the same in whichever runtime you work.
What’s missing right now
Section titled “What’s missing right now”This base honestly has unfinished parts. The manual and landing are public, but the template itself is a commercial license model, and the payment flow and buyer verification aren’t wired up yet. COMMERCIAL-LICENSE.md is still a draft and hasn’t passed lawyer review.
There are empty spots on the content side too. Right after a fork, _workspace/ is just a skeleton. What this manual teaches is the procedure by which that empty skeleton fills in through the 4-Tier flow, not a sample manuscript shipped inside book-studio.
What this manual does not assume
Section titled “What this manual does not assume”OS-by-OS install branches, a git intro, markdown syntax, how to use a publishing platform — none of that is here. If you need it, see the beginner manual. It’s the same base told at a different pace.
The next section is architecture. It covers what outputs the 4 Tiers hand off through and how several books are isolated by folder, with two diagrams.