Autonovel

Autonovel — pipeline entry point

This skill is the front door to the autonovel novel-writing pipeline. The user invokes it when they want to start a new novel, resume an existing one, or just understand what's available.

When to use this skill

• User says "let's write a novel about X" or "let's continue the novel"
• User mentions autonovel, the novel pipeline, or asks about phases
• User wants to see status of an in-progress novel

Prerequisites

The user must set AUTONOVEL_WORKSPACE to a writable directory (e.g. ~/novels). Each novel lives in $AUTONOVEL_WORKSPACE/$NOVEL_SLUG/ with its own state.json.

Check the workspace:

!`echo "AUTONOVEL_WORKSPACE=$AUTONOVEL_WORKSPACE" && test -d "$AUTONOVEL_WORKSPACE" && ls -la "$AUTONOVEL_WORKSPACE" || echo "workspace not set or missing"`

If AUTONOVEL_WORKSPACE is unset or the directory is missing, ask the user to set it before continuing. Default suggestion: ~/novels.

Workflow

Step 1 — Identify the novel

Ask the user: which novel? Options:

• Existing novel (list subdirs of $AUTONOVEL_WORKSPACE)
• New novel (ask for a short slug, e.g. "house-of-bells")

For a NEW novel:

1. Ask the user for a one-paragraph seed concept.

• 1b. If the novel is a commission (user pastes a subscriber request), run

references/commission-intake.md instead of asking for a seed paragraph — the approved pitch becomes the seed.

1. Create $AUTONOVEL_WORKSPACE/$NOVEL_SLUG/.
2. Copy ${HERMES_SKILL_DIR}/templates/seed.txt.template to $AUTONOVEL_WORKSPACE/$NOVEL_SLUG/seed.txt, fill in the seed.
3. Copy ${HERMES_SKILL_DIR}/templates/state.json.template to $AUTONOVEL_WORKSPACE/$NOVEL_SLUG/state.json, fill in the novel_slug.
4. Initialize a git repo in the novel directory so revisions are tracked.

Step 2 — Read state.json and dispatch

Read $AUTONOVEL_WORKSPACE/$NOVEL_SLUG/state.json. Look at phase:

• "foundation" → invoke autonovel-foundation skill
• "drafting" → invoke autonovel-drafting skill
• "revision" → invoke autonovel-revision skill
• "consistency" → invoke autonovel-consistency-pass skill
• "export" → invoke autonovel-export skill
• "done" → report status; user can re-enter via autonovel-prose-review for polish

For user-triggered prose review at any time, the user can directly invoke autonovel-prose-review regardless of state.

Step 3 — Show status if requested

If the user just asks "what's the status?", read state.json and report: current phase, iteration, chapter count, current threshold per dimension.

Reference docs

• references/pipeline-overview.md — high-level pipeline diagram and phase flow
• references/craft.md — craft principles (worldbuilding, character, prose) excerpted from upstream CRAFT.md
• references/anti-slop.md — list of patterns to avoid; informs every generation step
• references/ratings.md — the AO3 rating scale (G/T/M/E) and its enforcement rule
• references/commission-intake.md — subscriber request → approved content profile
• genres/ — genre reference packs loaded by profile tags (see genres/README.md)
• genres/AUTHORING.md — content-only pack contract; use through the manual

autonovel-genre-pack workflow, never during novel generation

Library utilities

The lib/ directory bundles:

• slop_detect.py — mechanical slop detection: banned words, AI tells, v2

metrics, genre tiers, and find_spans for exact finding locations (no LLM)

• slop_ngrams.py — manuscript-wide repeated-phrase / ending-overlap /

opener-dominance scanner (no LLM)

• grading.py — sentence-grading parser + deterministic acceptance gate
• story_state.py — story-state ledger: chapter-stamped facts, foreshadow

registry, STORY STATE query blocks, Mermaid relationship graphs

• namegen.py + namegen_data/ — name generation from 28 style packs with

banned-attractor filtering and the check_cast distinctiveness gate

• state.py — state.json read/write + adaptive threshold computation

(thresholds are trend signals only; acceptance is gate-based)

• chunk_text.py — token-aware text windowing
• genre_registry.py + genre_resolver.py — manifest discovery, deterministic

profile resolution, and bounded task-specific packet compilation

• genre_patterns.py — selected pack pattern loading with provenance
• genre_pack.py — manual scaffold, validation, and context inspection commands

These are reused across phase skills. Invoke via python ${HERMES_SKILL_DIR}/lib/.py or import them into ad-hoc scripts during execution.

Workspace conventions

Each novel directory contains:

$NOVEL_SLUG/
├── state.json                 # pipeline state
├── seed.txt                   # original concept
├── world.md                   # phase 1 output
├── characters.md              # phase 1 output
├── outline.md                 # phase 1 output
├── canon.md                   # phase 1 output
├── voice.md                   # phase 1 output
├── genre_context/             # deterministic genre resolution + packets
│   ├── resolved.json          # selections, warnings, sources, budgets, fingerprints
│   ├── story-contract.md      # persistent story-specific genre promises
│   └── compiled/
│       ├── contract.md
│       ├── world.md           # {GENRE_WORLD_BLOCK}
│       ├── characters.md      # {GENRE_CHARACTER_BLOCK}
│       ├── outline.md         # {GENRE_OUTLINE_BLOCK}
│       ├── drafting.md        # {GENRE_DRAFTING_BLOCK}
│       ├── evaluation-chapter.md     # {GENRE_EVALUATION_BLOCK}
│       ├── evaluation-manuscript.md  # {GENRE_EVALUATION_BLOCK}
│       └── patterns.json
├── chapters/
│   └── ch_01.md, ch_02.md, ...
├── story_state/               # coherence ledger (phase 2+)
│   ├── facts.json             # entities + chapter-stamped facts + foreshadow
│   └── graph_act_N.md         # Mermaid relationship diagrams (for humans)
├── consistency_state/
│   └── rolling_summary.md     # per-chapter summaries, appended live at drafting
├── critique/                  # phase 3 outputs
│   ├── adversarial_NN.md
│   ├── reader_panel_NN.md
│   ├── consistency.md
│   ├── slop_report_.md # user-triggered autonovel-grade reports
│   └── prose/                 # per-chapter or per-scope notes from prose-review skill
└── export/                    # phase 4 outputs (PDF, ePub, etc.)

State is persisted to git via the novel-directory's own repo. Commit after every meaningful step.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: DukeTwoCan
• Source: DukeTwoCan/autonovel-agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.