Book Pipeline

Book Pipeline — long-form AI fiction production

Turn many small story ideas into publishable short novels or audiobooks through a delegated, iterative pipeline. Built on narrative v2 (scene craft + project mode) and narrative-revisor (separate critic).

When to use

• User wants to write a novel / short novel / audiobook with AI assistance.
• User is building or refining a book-writer profile.
• User has multiple story ideas to expand and publish.
• Any request combining "draft fiction" + "review loop" + "publish/export."

Architecture

Operating principle: do NOT do all domain work in one profile. Orchestrator (Senna) routes; book-writer owns manuscript state. Lanes use delegate_task when multi-domain.

User phrases (hard protocol):

• "use the team" → parallel delegate_task to specialists; do not solo multi-domain work from the orchestrator.
• "don't do all the work" / "as needed" → propose lanes; batch go-ahead when ambiguous; never expand a skill test into a full pipeline build unprompted.
• "after get all back then report" → hold partials; one consolidated brief when every lane returns (status table + conflicts + recommendation). Never drip five half-reports.

Report protocol: multi-lane fan-out → wait for all → one consolidated brief. See references/orchestrator-fanout.md.

Locked decisions: references/locked-decisions.md — Darwin-only; stdlib assembler canonical; Kokoro→Findaway. Build lessons: references/session-lessons-pipeline-build.md.

End-to-end loop (wired)

0 INIT     init_manuscript.py → scaffold tree
1 CONCEPT  user + book-writer → concept.md + tone
2 INTEL    research lane (optional but recommended pre-draft)
3 PLAN     character-sheet, plot-ledger, foreshadow-bank, worldbuilding, voice-profile
4 DRAFT    narrative skill → chapters/chNN.md  status:drafted
5 REVIEW   narrative-revisor → reviews/*.json   (fresh context)
6 REVISE   apply issues only → re-review until pass
7 GATE     check_manuscript.py + chapter checklist → status:done
8 EXPORT   python -m manuscript_pipeline.cli --book <name> --out out
9 PUBLISH  business lane (KDP Select → D2D; audio via Findaway)

Stage contracts

Commands

# Scaffold (from repo or skill scripts dir)
python ~/.hermes/profiles/senna/skills/writing-and-review/book-pipeline/scripts/init_manuscript.py my-title \
  --root ~/book-writer-pipeline/manuscript --title "My Title"

# Mechanical scan
python ~/.hermes/profiles/senna/skills/narrative/scripts/check_manuscript.py \
  ~/book-writer-pipeline/manuscript/my-title

# Export (canonical zero-dep CLI)
cd ~/book-writer-pipeline
python -m manuscript_pipeline.cli --book my-title --out out

Windows Darwin stack: book-writer-pipeline/windows/README.md + start-darwin-server.bat.

Manuscript folder convention

Copy templates/manuscript-layout.md or run scripts/init_manuscript.py.

manuscript/{project}/
  manuscript.yaml
  concept.md
  canon.md
  character-sheet.md
  plot-ledger.md
  foreshadow-bank.md
  worldbuilding.md
  voice-profile.md
  chapters/ch01.md …
  outlines/synopsis.md
  reviews/                 # revisor JSON
  exports/                 # CLI outputs (optional mirror)

Model note (locked: Darwin-only)

Local: Windows your GPU / 32GB RAM. Only local-36B-Opus APEX I-Mini.
llama.cpp llama-server + --n-cpu-moe. Sequential plan/draft/review — never concurrent loads.
No speculative decoding on MoE. Inject ledger slices, never full manuscript.
TTS: Kokoro-82M → Findaway/Spotify. ACX rejects third-party TTS.
API fallback: OpenRouter DeepSeek V3.2 or Qwen3.6-Plus.

Skills map

Pitfalls

• Consistency drift without ledger enforcement
• One profile drafting AND reviewing its own prose
• Context overflow from full-manuscript dumps
• Stability trap only at scene grain (must also run book-end)
• Market research after writing instead of before
• Dual local models when only Darwin is available
• Parallel fan-out duplicates artifacts. Two CLIs or model briefs → mark CANONICAL vs DEPRECATED immediately; update locked-decisions.md. Prefer stdlib assembler unless user asks for covers/print quality.
• Subagents invent skills / patch SKILL.md unprompted. After fan-out, audit the skills tree; keep only intentional patches.
• Scaffold + str.format. YAML braces break Python .format(). Use __PLACEHOLDER__ + .replace() in scripts/init_manuscript.py — never {slug} inside YAML templates.
• Profile describe CLI: hermes profile describe <name> --text "…" (not --description).
• macOS vs Windows host: ship book-writer-pipeline/windows/* for the your GPU; do not claim Darwin is live until that box starts llama-server.
• Expanding a narrative skill test into architecture/profile work without being asked.

Support files

• references/locked-decisions.md
• references/orchestrator-fanout.md
• references/style-revision-playbook.md
• references/session-lessons-pipeline-build.md — build pitfalls + smoke sequence
• references/zero-dep-export.md
• references/build-export-scaffold.md (deprecated)
• templates/manuscript-layout.md
• scripts/init_manuscript.py
• Companion: narrative/scripts/check_manuscript.py; book-writer-pipeline/windows/*

Verify

• Manuscript tree has state files before chapter 2
• Each chapter has reviewer pass JSON before assembly
• check_manuscript.py clean of blockers before export
• Export produces openable EPUB
• Multi-lane work ends in one synthesis report
• After multi-agent design: one canonical CLI, one model story, skills tree audited