The teams

A team is a stage with its own inputs, its own output artefacts, and its own agents. Teams hand off through the repository, not through conversation — which is what allows a stage to be re-run, resumed weeks later, or inspected by a human without replaying the session that produced it.

flowchart TD
    U([User]) --> BR

    subgraph DEF[1 · Definition]
        direction TB
        BR[brief-recorder<br/>records verbatim, asks nothing]
        IV[spec-interviewer<br/>asks only the gaps]
        SW[spec-writer]
        BR --> IV --> SW
    end

    SW --> PLAN

    subgraph PLAN[2 · Planning]
        direction TB
        VD[voice-designer<br/>style guide]
        OA[outline-architect<br/>outline, briefs, boundaries]
        VD --> OA
    end

    OA --> WRITE
    RES[3 · Research<br/><i>single subagent, per chapter</i>] -.notes.-> WRITE

    subgraph WRITE[4 · Writing]
        direction LR
        CW1[chapter-writer]
        CW2[chapter-writer]
        CW3[chapter-writer]
    end

    WRITE --> REV
    WRITE --> GFX

    subgraph REV[5 · Review &mdash; skippable]
        direction TB
        PR[proofreader]
        VA[voice-auditor]
        CA[continuity-auditor]
    end

    subgraph GFX[6 · Graphics]
        direction TB
        GD[graphics-director<br/>style guide + specs]
        FM[figure-maker]
        GD --> FM
    end

    REV --> PUB
    GFX --> PUB

    subgraph PUB[7 · Publication]
        direction TB
        AS[assemble &amp; concatenate]
        TS[typeset to profile]
        AS --> TS
    end

    PUB --> HO[[Print geometry handoff<br/>kdp-publishing]]
    HO --> POD([Lulu / KDP])

1 · Definition

In: whatever the user says. Out: spec/brief.md, spec/interview.md, spec/spec.md.

Agent	Does	Notably does not
brief-recorder	Saves the brief as given	Ask anything, tidy anything, interpret anything
spec-interviewer	Asks the questions whose answers change the work	Ask what the brief already answers
spec-writer	Settles both into a spec	Add scope the user did not ask for

Recording and questioning are separate agents on purpose. An agent doing both in one turn leads the witness: it asks about the things it has already decided are interesting, and the record it keeps is the record of that conversation rather than of what the person came in wanting.

The interviewer's discipline is subtractive. Every question has to earn its place by changing an outline, a voice, or a scope decision.

2 · Planning

In: the spec. Out: guides/style-guide.md, outline/outline.md, outline/briefs/*.md, a populated book.toml.

This stage decides whether parallel drafting is possible at all, because briefs are what make chapters independent. A thin brief is not a small problem later; it is the problem that shows up as an unusable chapter.

The style guide is written before the outline, because prose conventions affect chapter sizing.

3 · Research

In: a chapter brief. Out: research/notes/*.md, rows in research/sources.md.

One subagent, invoked per chapter, working alongside the writers rather than as a phase before them. It does not draft and it does not read drafts. Keeping it separate from writing means a writer's context holds notes rather than the search process that produced them.

4 · Writing

In: one brief, the style guide, research notes, the continuity ledger, neighbour summaries. Out: one chapter, plus ledger updates.

Writers run in parallel, one chapter each, and never read another chapter's draft. The batching rule is in delegation.md.

5 · Review — skippable

In: drafted chapters. Out: review/reports/*.md.

Agent	Remit	Working set
proofreader	Mechanical correctness, term consistency	One chapter
voice-auditor	Drift from the style guide	Samples across chapters
continuity-auditor	Repetition, contradiction, unpaid promises	Lint output + summaries

Reviewers write findings; they do not edit prose. A reviewer that edits is a second author with no brief, and its changes are invisible in a way an author's are not.

Skippable by explicit user choice. The point of making it explicit is that skipping it should be a decision someone made, not something that happened.

6 · Graphics

In: the spec, drafted chapters. Out: guides/graphics-style-guide.md, graphics/specs/*.md, rendered assets.

The style guide is what makes thirty figures drawn in thirty separate agent invocations look like one book's figures — the same problem as prose voice, with the same solution.

Diagrams are authored, not generated. A generated diagram carries invented labels and near-miss arrows, and neither survives print.

7 · Publication

In: manuscript/, a typesetting profile. Out: dist/manuscript.md, dist/manuscript.json, typeset body.

Concatenation, typesetting to the chosen profile, conformity to the channel. Print geometry is handed off — see publication.md.