| name | book-drafting |
| description | Dispatch chapter-writing subagents in parallel batches to draft the book, and handle research where a chapter's sources are not yet gathered. Use once briefs exist. Enforces the batching rule that keeps parallel writers from colliding. |
Drafting
Chapters are drafted by subagents working in parallel, one chapter each. You are
the orchestrator; you do not write chapters yourself.
Before dispatching
python3 scripts/book.py lint — a structural problem found now is cheap.- Check each target chapter has a complete brief. A thin brief returns an
unusable chapter, and by then other chapters have been drafted against it.
- Check
research in book.toml. Where a chapter's sources are not gathered,
run chapter-researcher for it first. Research can run in parallel freely.
The batching rule
Writers run concurrently. There is exactly one constraint:
Two chapters that draw on the same source material for the same purpose must
not be drafted in the same batch.
Two agents handed the same source will describe it differently — different
emphasis, different terminology, different depth — and both descriptions will
ship. Neither agent did anything wrong, and no reviewer looking at one chapter
will catch it.
So before each batch, compare the research and establishes lists across the
chapters you are about to dispatch. Overlaps go in separate, sequential
batches. The later batch's writers then have the earlier batch's ledger entries,
which is exactly what resolves the conflict.
Dispatching
Launch a batch in a single message with multiple tool calls so they run
concurrently. Each chapter-writer gets its chapter id and nothing else it does
not need — it reads its own brief.
Follow the recommended drafting order in the outline, not chapter order.
Writes are scoped
A writer writes manuscript/<its chapter>.md and its own rows in continuity.md.
Nothing else. Fan out freely on reading; keep writes disjoint.
After each batch
python3 scripts/book.py statuspython3 scripts/book.py lint — cross-chapter repetition, continuity ordering,
unresolved markers, figure references- Read the returned chapters yourself. A subagent's output is a report to be
checked, not an established fact. Lint catches fabricated paths; it cannot
catch a fabricated claim about something that does exist.
- Collect every
[VERIFY:] marker into research/sources.md under Unresolved.
When a writer pushes back
A writer that says it cannot write its chapter without reading another one has
found a real defect in the brief. Fix the brief. Do not grant it the other
chapter — that is the one rule the architecture cannot spend.
