Kinoshita Narrative Pipeline

A disciplined multi-agent pipeline for drafting and editing fiction in the Deckline Cycle — the in-world novelettes set in the KN-86 Deckline universe. Adapted from the Amber Circuit production run; generalized for any volume in the cycle. Operates in two modes (WRITE / EDIT) over the same agent topology.

The fiction is load-bearing marketing. Every Deckline Cycle volume is a vehicle for exposing readers to the KN-86’s capability cartridges, runtime systems (mission board, CIPHER voice, Universal Deck State), and the device’s sensory signature (amber-on-black 80×25, OLED CIPHER-LINE strip, YM2149 PSG audio). The reader should finish each book curious about the device — what it is, how it works, what cartridges they’d run. Pipeline gates enforce this.

When to use

• User asks to write a new piece in the Deckline Cycle (novelette ~12–20k, novella 20–40k, short story 2–7k).
• User asks to edit an existing v_N draft for canon alignment, voice consistency, or extensive rewrite.
• User mentions Variance Analysis, The Amber Circuit, The Lisp Machine, The Operator’s Manual, The Inheritance, Amber and Dust, or any anthology piece.
• User requests “canon check,” “alignment with [other book],” “cross-cycle continuity,” or “make this hit harder.”

Mode dispatch — the orchestrator’s first decision

Before spawning any agent, decide which mode you’re in.

Trigger	Mode	Pipeline
”write a new novelette,” “draft a story,” no manuscript file exists	WRITE	Research → PM (outline) → Author (chapters) → Editor (developmental) → Copyeditor (line) → Production
”extensively edit,” “align with,” “canon-check,” manuscript file exists at stories/<Title>-v3.md	EDIT	Research → Alignment Audit (PM) → Editor (developmental rewrite) → Copyeditor (line) → Production

If ambiguous, ask the user. Don’t guess.

The Variance Analysis ask as of 2026-04-26 is EDIT mode — the v3 draft exists at stories/Variance-Analysis-v3.md and is being aligned with the post-Amber-Circuit canon.

The Kinoshita World — single source of truth

Master worldbook (canonical, current): Kinoshita World in Notion

Local mirror (may lag Notion): docs/marketing/narrative/worldbook.md

When canon conflicts: Notion master wins for fiction-only lore (factions, characters, ledgers, AetherNet rules, CIPHER voice arc); engineering documents (docs/software/, docs/device/, CLAUDE.md Canonical Hardware Specification) win for hardware/runtime facts. The Research agent reconciles both into the dossier; downstream agents trust the dossier.

Cross-cycle anchor index: see reference/kinoshita-world-canon.md for known cross-cycle phrase rhymes (“autonomous output generation” linking Variance Analysis ↔ Lisp Machine, etc.), the Edgeware/Kinoshita corporate lineage, and the canonical CIPHER voice rules.

The marketing obligation — why this skill is different

A Deckline Cycle book is not standalone fiction. It is a brochure for a fictional console that actually exists. Every book ships with these enforced obligations, in addition to the literary obligations of plot/character/voice:

1. Capability coverage. The book must meaningfully use a stated set of capability cartridges (per-volume coverage table — see synopsis files). Each cartridge must be used in action with sensory consequence, not name-dropped in dialogue. The PM/Alignment-Audit agent maintains a coverage table; the Editor enforces it.
2. Runtime features surfaced. The mission board, CIPHER voice (CIPHER-LINE OLED only, never main grid), Universal Deck State, reputation tier gating, AetherNet stealth properties, Black ICE, and TradEcon/synthetic-identity economics must each appear at least once across any given book or be explicitly waived in the synopsis.
3. Sensory fingerprint. Amber on black. 80×25 grid. CIPHER fragments scrolling on the OLED strip. YM2149 phase tones. The reader should be able to picture the device after reading.
4. Curiosity hook. The book should raise more questions than it answers. A reader closing the back cover should want to know: what would I run on this thing? The marketing-strategist follow-up step (see Production) confirms the back-cover copy plants this hook.

See reference/kn86-feature-marketing.md for the full feature surface and per-cartridge marketing angles.

Agent topology

You orchestrate; the agents execute. Agents share NO memory — every brief is self-contained with file paths to read and rules to follow. Spawn each via the Agent tool with the relevant brief from briefs/.

WRITE mode:
  RESEARCH → PM (outline) → AUTHOR (chapters, chunked) → EDITOR (developmental) → COPYEDITOR (line) → PRODUCTION

EDIT mode:
  RESEARCH → ALIGNMENT AUDIT (PM) → EDITOR (developmental rewrite) → COPYEDITOR (line) → PRODUCTION

Agent	Role	Brief	Subagent type
Research	Read worldbook + cycle + relevant ADRs/specs; produce dossier	briefs/research-agent.md	Explore
PM (write)	Outline + character bible + style brief from dossier	briefs/pm-agent.md	general-purpose
PM (edit / Alignment Audit)	Read existing draft; produce gap analysis vs. canon + cross-cycle + marketing obligations	briefs/pm-agent.md	general-purpose
Author	Draft prose chapters chunked by 2–3 chapters per call	briefs/author-agent.md	general-purpose
Editor	Developmental pass; rewrite passages directly; not just notes	briefs/editor-agent.md	general-purpose
Copyeditor	Line-level polish; assemble manuscript	briefs/copyeditor-agent.md	general-purpose
Production	Format .docx + .epub + cover binding + metadata; orchestrator runs this directly	briefs/production-orchestrator.md	(you)

Co-working note. The agents are sequential by default — output of phase N is input to phase N+1. The orchestrator (you) is the only memory of the whole production. Read each phase’s output before dispatching the next; bounce work back via SendMessage rather than re-spawning when corrections are minor.

Filesystem conventions (Deckline Cycle)

This skill writes into the established narrative tree at docs/marketing/narrative/. Do not invent new paths.

Artifact	Path
Local worldbook mirror	docs/marketing/narrative/worldbook.md
Working drafts (current published version)	docs/marketing/narrative/stories/<Title>-v3.md
In-flight revisions	docs/marketing/narrative/stories/<Title>-v4-draft1.md, <Title>-v4-draft2.md, etc.
Final cuts	docs/marketing/narrative/stories/<Title>-v4.md (overwrite v3 only after Josh approves)
Synopses (canon obligation tables live here)	docs/marketing/narrative/synopses/<Title>-Synopsis.md
Changelogs (per draft, per phase)	docs/marketing/narrative/changelogs/<Title>-<phase>-changelog.md
Production output	docs/marketing/narrative/build/<title-slug>/{body.md,metadata.yaml,style.css,cover.png,back-cover.png,<Title>.epub,<Title>.kpf}
Per-run scratch (research dossiers, outlines, briefs)	docs/marketing/narrative/.work/<title-slug>-<run-id>/

The .work/ directory is where the orchestrator stages dossier, outline, character-bible, style-brief, editor-memo, alignment-audit between agent phases. Gitignore-able. Don’t commit dossiers; commit the manuscript and the changelog.

WRITE mode — six phases

Phase 0 — Scope & clarification (orchestrator)

Before any agent. Use AskUserQuestion for anything underspecified.

Parameter	Default if unspecified	Example
Length	Novelette ~14k	”novelette,” “novella,” “short story”
Structure	6 chapters × 2,300 words	”5 chapters,” “8 chapters”
POV/Tense	Close third, past tense, single POV	”first-person present”
Voice reference	Per-volume voice DNA (see reference/voice-dna.md); ask if new volume	”Gibson + Vonnegut + Morgan,” “DeLillo institutional”
Volume in cycle	Ask	”next book after Variance Analysis”
Capability coverage	Ask; minimum 8 cartridges; spread across the arc	”must use ICE BREAKER, Black Ledger, CIPHER Garden…”
Deliverables	.epub + .docx + body.md + cover artwork brief + synopsis	”audiobook-ready epub”

Phase 1 — Research (Explore subagent)

Spawn an Explore subagent with briefs/research-agent.md. It produces a dossier at .work/<run-id>/dossier.md. Review the dossier before proceeding — it must cover every obligation cartridge with sensory detail liftable into prose.

Phase 2 — PM outline (general-purpose subagent)

Spawn the PM with briefs/pm-agent.md in write mode. It produces:

• .work/<run-id>/synopsis.md — spoiler-full prose synopsis (~700 words)
• .work/<run-id>/character-bible.md
• .work/<run-id>/outline.md — chapter-by-chapter, scene-by-scene, with the Capability Distribution Map
• .work/<run-id>/style-brief.md — voice rules, do/don’t lists, pastiche calibration paragraph

Review gate: does the outline earn its ending? Every cartridge in the Distribution Map? Word budget sums correctly? Is the marketing curiosity hook planted in the opening chapter and paid off in the final chapter? If any answer is no, send corrections via SendMessage before proceeding.

Phase 3 — Author (general-purpose subagent, chunked)

Spawn the Author with briefs/author-agent.md. For novelettes, split: chapters 1–3, then chapters 4–6 (or however many — the second invocation MUST read the first to maintain continuity). For novellas (>20k), chunk 2–3 chapters per call.

Review gate after each chunk: word counts within ±15%? Voice consistent with the style brief’s pastiche paragraph? Cartridges from the Distribution Map appearing in scene with sensory consequence? Seeds planted for later chapters? If summary instead of scene, re-brief the agent specifically.

Phase 4 — Editor developmental pass (general-purpose subagent)

Spawn the Editor with briefs/editor-agent.md. It reads everything, identifies structural/voice/canon problems, and rewrites affected passages directly in the chapter files. It does not just leave notes. Output: revised chapters + an editor-memo at .work/<run-id>/editor-memo.md.

Review gate: does the ending land? Voice consistent across all chapters? CIPHER voice arc shaped (procedural early → familiar late)? Every capability used in action? Word count still on target?

Phase 5 — Copyeditor (general-purpose subagent)

Spawn the Copyeditor with briefs/copyeditor-agent.md. Line-level pass; assemble manuscript at .work/<run-id>/manuscript.md. Then promote to stories/<Title>-v_next-draft1.md for the changelog record.

Phase 6 — Production (orchestrator, no agent)

Run the production steps directly per briefs/production-orchestrator.md. Generate body.md, metadata.yaml, build .epub (with epubcheck validation) and .kpf for KDP, write the changelog, commit.

EDIT mode — five phases

EDIT mode collapses the PM (outline) and Author (drafting) phases into a single Alignment Audit phase that compares the existing draft to canon, then dispatches the developmental Editor to rewrite. Use this when a manuscript exists at v3 (or any v_N) and needs extensive revision rather than re-drafting.

Phase 0 — Confirm scope (orchestrator)

Ask: which book? Which canon-alignment targets? Cross-cycle anchors to preserve? Word-count tolerance for cuts/additions? Voice direction (preserve or shift)?

Phase 1 — Research (Explore subagent)

Same as WRITE mode, but the dossier prompt asks specifically for canon as it stands right now (post-most-recent-changes) and cross-cycle continuity points the edit must preserve. The Research agent reads:

• The Notion worldbook (master)
• docs/marketing/narrative/worldbook.md (local)
• The other v3 stories (for cross-cycle anchors)
• The relevant synopsis (canon obligation table)
• The most recent changelog entries (what changed in the last revision)

Phase 2 — Alignment Audit (PM, edit mode)

Spawn the PM with briefs/pm-agent.md in edit mode. It produces an alignment audit at .work/<run-id>/alignment-audit.md with three sections:

1. Canon deltas — places where the draft contradicts current canon (factual, ontological, terminological).
2. Cross-cycle deltas — places where the draft contradicts other v3 books (corporate names, character ages, AetherNet rules, CIPHER behavior, ICE class taxonomy).
3. Marketing deltas — capability cartridges from the synopsis coverage table that appear only in name-drop, not in action; runtime features absent that should be present.

Output is a prioritized list of rewrites the Editor will execute, mapped to specific chapter / paragraph / line ranges in the draft.

Review gate: is the audit specific enough that the Editor can act on it without inventing the fix? If items say “this scene needs more sensory detail,” send back for specific what and where.

Phase 3 — Editor developmental rewrite (general-purpose subagent)

Spawn the Editor with briefs/editor-agent.md in edit mode. It executes the alignment audit’s rewrites directly in the manuscript file (working in .work/<run-id>/draft-edited.md, then promoted). Same rule as write mode: rewrite passages, don’t just leave notes.

Review gate: spot-check the rewritten passages against the audit. Did the Editor address every flagged item? Voice still on-key for this volume? No new canon contradictions introduced?

Phase 4 — Copyeditor (general-purpose subagent)

Same as WRITE mode phase 5.

Phase 5 — Production (orchestrator)

Same as WRITE mode phase 6, plus: write a structured changelog under changelogs/<Title>-rev<N>-changelog.md enumerating canon deltas applied, cross-cycle alignments, and marketing deltas closed. Reference the changelog format in existing files (e.g., changelogs/The-Amber-Circuit-rev1-changelog.md).

Failure modes & recovery

Symptom	Diagnosis	Fix
Author returns chapters that summarize plot instead of dramatizing scenes	Brief was too abstract	Re-brief specifically: “Write the scene at [location]. Open with [image]. Render [character action]. Close on [turn].” Cite the outline beat verbatim.
Voice drifts mid-book (chapter 4 sounds different from chapter 1)	Pastiche calibration not load-bearing	SendMessage to Author with chapter 1 paragraph as calibration target
Capability cartridge appears only in dialogue (“they used Black Ledger”)	Marketing obligation violated	SendMessage to Editor: “Rewrite [scene] so the character physically uses Black Ledger. Show the cell stack, the audit graph, the CV redaction. Sensory + consequence.”
Editor returns a memo but doesn’t change the chapters	Brief did not enforce direct rewrite	Re-spawn Editor with explicit: “Rewrite the affected passages in the chapter files. The memo summarizes what you did, not what should be done.”
Word count >20% off target	Author / Editor freestyled length	Identify which scenes are short / long; re-brief with explicit per-scene word allocation
Cross-cycle contradiction (e.g., character age different from another book)	Research dossier missed it, or canon shifted between runs	Patch the dossier; alert Editor with the specific contradiction and the canonical answer
Marketing curiosity hook absent from back cover	Production step skipped the marketing copy review	Use marketing-strategist skill to draft jacket copy that plants the device curiosity hook
Notion worldbook updated mid-run	Master canon shifted	Pause the active phase; re-fetch worldbook; patch dossier; resume

Adjacent skills to load

The orchestrator should load these alongside this skill, by context:

• task-management — every Deckline Cycle revision is a Notion-tracked task. Open it before dispatching agents; close it after Production with the changelog link.
• brainstorming — before WRITE mode if the volume’s premise isn’t fully formed.
• copywriting — for back-cover copy, jacket flap, marketing dispatches that surface the book.
• marketing-strategist — for the curiosity-hook validation in Production.
• copy-editing — provides the line-edit pass discipline; the Copyeditor brief references its conventions.
• markdown-to-kindle-epub — used in Production to build .epub for KDP.
• git-flow — manuscript revisions get branches and PRs like code; commit the changelog separately.

What this skill does NOT do

• Write standalone fiction unrelated to the KN-86. Use short-story-writing for that.
• Generate non-fiction marketing content (blog posts, dispatches, social). Use copywriting / social-content / content-strategy.
• Author the in-fiction operator manuals (the player-facing texts at docs/software/cartridges/operator-manuals/). Those are diegetic instruction manuals, written under different conventions; ask the user to spawn a separate agent.
• Replace the orchestrator’s judgment. Review every phase output. Don’t auto-merge.