book-writing

A Claude Code plugin for writing a full-length reference or instructional book with teams of subagents — organised so that no single agent is ever given a task large enough to degrade.

/plugin install book-writing@danielrosehill

What it is for

Producing a print-on-demand paperback — Lulu, KDP Create, or equivalent — that carries educational, instructional or reference material which did not exist in that form before. Manuals, field guides, technical references, the documentation of a system nobody has written down.

It is an agentic research pattern in which the book is the delivery mechanism. It is deliberately not built for producing books at volume: the pipeline gates on a definition stage a volume generator would make optional, it binds writers to evidence they can reach, and its unit of value is one book worth printing.

The problem it solves

One agent cannot write a book. Not because a manuscript will not fit in a context window — increasingly it does — but because quality degrades long before the window fills. A single agent carrying a whole book restates what it said earlier, drifts in register, echoes its own chapter openings, and starts producing text that is locally fluent and globally empty.

So the architecture is context management by delegation. Every writing task is one chapter, from one brief.

That has a cost, and the plugin is mostly machinery for paying it: agents that do not read each other's work cannot stay consistent by reading, so consistency is carried by documents instead — a style guide, per-chapter briefs, a continuity ledger. That trade is the whole pattern.

The teams

flowchart LR
    U([User]) --> D[Definition]
    D --> P[Planning]
    P --> W[Writing]
    R[Research] -.-> W
    W --> RV[Review]
    W --> G[Graphics]
    RV --> PB[Publication]
    G --> PB
    PB --> H[[kdp-publishing]]

Team	Agents
Definition	brief-recorder, spec-interviewer, spec-writer
Planning	voice-designer, outline-architect
Research	chapter-researcher
Writing	chapter-writer × n, in parallel
Review	proofreader, voice-auditor, continuity-auditor
Graphics	graphics-director, figure-maker
Publication	assembly scripts, then handoff

Teams hand off through the repository, not through conversation — which is what lets a stage be re-run, resumed weeks later, or inspected without replaying the session that produced it.

Commands

/book-new	Create the project workspace from the template
/book-brief	Record the brief exactly as given, asking nothing
/book-interview	Ask only the gaps that change the work
/book-spec	Settle the spec
/book-style	Write the style guide
/book-outline	Design the structure and write per-chapter briefs
/book-research	Gather a chapter's sources
/book-draft	Draft chapters with parallel writers
/book-review	Proofread, voice audit, continuity audit
/book-graphics	Visual language and figures
/book-assemble	Assemble and hand off to print geometry
/book-status	What is drafted, what is blocking

Three rules that carry most of the value

Recording and interviewing are separate agents. brief-recorder asks nothing and interprets nothing. An agent that records and questions in one turn leads the witness, and the user's original framing is unrecoverable once paraphrased.

Writers never read other chapters' drafts. Text in context is text to imitate: a writer that has read chapter three reuses its transitions, borrows its examples and inherits whatever drift it had already accumulated. Each brief carries neighbour summaries as the controlled substitute.

Two chapters drawing on the same source do not draft concurrently. Two agents given the same source describe it differently, and both descriptions ship. Neither agent did anything wrong and no per-chapter reviewer will catch it.

What it does not do

It does not typeset your book. Trim, gutter, spine width, cover wrap and pre-flight belong to kdp-publishing, which builds the cover last from the interior's measured extent rather than from an estimated page count. /book-assemble produces dist/manuscript.md and a JSON sidecar and hands off there.

The division runs both ways: that plugin states it does not write your book, and this one does not typeset it. Typesetting and authorship have different failure modes, and tools that bundle them tend to be good at neither.

It does not verify facts. book.py lint checks that a source is cited, never that it says what the chapter claims. A chapter can pass every mechanical gate while making a confident, specific, invented claim. Someone has to read the book.

The workspace

Each book lives in its own repository, instantiated from Claude-Book-Workspace-Template. That repo is canonical; template/ here is a byte-identical vendored copy so /book-new can scaffold locally when gh is unavailable.

./scripts/sync-from-template.sh --check   # report drift
./scripts/sync-from-template.sh           # refresh the vendored copy

A workspace works standalone. python3 scripts/book.py status | lint | assemble run with this plugin uninstalled, because a book outlives a plugin version.

The pattern

Specified and documented at Agentic-Book-Writing. SPEC.md there is the source of truth; this plugin is an implementation of it and is wrong where the two disagree.

MIT.