Autonovel drafting

Autonovel — Phase 2 (Drafting)

Draft each chapter in sequence from the scene-level outline. Per-chapter loop: draft → mechanical gates → surgical rewrite loop → grading gate (see references/surgical-rewrite.md).

When to use this skill

• State.json shows "phase": "drafting"
• Foundation complete; outline.md has scene-level beats; world/characters/voice/canon all present
• chapters_drafted < chapters_total (or chapters_total is 0 and we need to count from outline)

Prerequisites

!`cd "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG" && ls -la outline.md world.md characters.md voice.md canon.md`
!`cat "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/state.json"`

If any foundation file is missing, hand back to autonovel-foundation.

Workflow

Step 1 — Resolve the next chapter

Read chapters_drafted + 1 as the next chapter number. New foundations already carry chapters_total. For a legacy drafting state where it is zero, count the validated outline headings for the loop bound; the first atomic checkpoint persists that derived count. Do not edit state.json by hand.

Step 2 — Draft loop

For each chapter index N from chapters_drafted + 1 to chapters_total:

a. Compile and build context for chapter N. At the start of every chapter, run the deterministic preflight before spending a model call:

python ${HERMES_SKILL_DIR}/lib/validation.py preflight "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG" --chapter N

Nonzero exit means the accepted foundation does not contain a safe drafting packet for this chapter. Report the returned issue codes and stop this chapter; the command does not mutate state. After preflight, re-run the task-specific resolver (unchanged inputs use its fingerprint cache):

python ${HERMES_SKILL_DIR}/../autonovel/lib/genre_resolver.py compile --state "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/state.json" --workspace "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG" --task drafting.chapter

Read:

• outline.md — get the scene-beat section for chapter N
• world.md, characters.md, voice.md, canon.md — full
• chapters/ch_{N-1:02d}.md if N > 1 — get the last 1000 words for continuity (prose texture and immediate scene state; NOT a source of facts — see the STORY STATE block below, which replaces re-reading older chapters for facts)
• references/chapter-prompt.md (THIS SKILL's reference dir) — get the chapter draft template
• genre_context/story-contract.md — the persistent story-specific genre promises
• genre_context/compiled/drafting.md — the bounded static drafting packet
• genre_context/compiled/evaluation-chapter.md — the profile-compliance and semantic genre-slop criteria used by Gate 4
• genre_context/resolved.json — selected-pack metadata and the generated pattern-bundle pointer used by the mechanical scan
• profile from state.json — rating, exclusions, and content tags to match against the outline coverage map; do not reopen source pack Markdown from Drafting
• STORY STATE block (if story_state/facts.json exists — see references/fact-extraction.md for how it's populated):

  !`python3 ${HERMES_SKILL_DIR}/lib/story_state.py query "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/story_state/facts.json" {N} <characters-in-this-beat>`
  

  <characters-in-this-beat> = the entity keys of characters named in chapter N's outline beat (space-separated CLI args). Include the returned facts and participating characters in {GENRE_DRAFTING_BLOCK} rather than duplicating a separate story-state block. Budget: this story-state portion must stay <= 1.5k tokens; this cap does not include or truncate the story contract or compiled drafting packet. If the query runs longer, narrow the subject list further to only the characters actually on-page in this beat rather than the full cast. If story_state/facts.json does not exist yet, skip this block — it is created on the first chapter's fact-extraction step (references/fact-extraction.md).

Assemble {GENRE_DRAFTING_BLOCK} in this order:

1. genre_context/story-contract.md.
2. genre_context/compiled/drafting.md.
3. The exact profile line: primary genre and primary engine, all positive tags, content tags, shape, and pitch.
4. The active rating's full rating definition from ${HERMES_SKILL_DIR}/../autonovel/references/ratings.md, followed by every exclusion as a literal NEVER include requirement.
5. This chapter's content-tag coverage assignments.
6. Current story-state facts and the participating characters for this chapter.

Insert that single block into references/chapter-prompt.md. The story contract and compiled packet are authoritative static sources; append only the chapter-specific requirements above. Never read craft.md, beats.md, slop.md, guardrails.md, or any other source pack Markdown directly here.

b. Generate the chapter. Fill the chapter-prompt template with the assembled context. Target word count from outline (typically 3000-5000 words per chapter). The agent does the LLM work directly using Hermes's loop.

c. Write to chapters/ch_{N:02d}.md.

Run the hard chapter validator immediately. A voice-calibration echo or invalid artifact must be fixed before the broader quality loop. The validator also enforces the outline's word-count target with a ±15% tolerance; a chapter.word_count issue requires expanding or tightening the chapter before any quality score or checkpoint. Put every scene-level insertion or cut needed for that repair into one patch document and apply it once with lib/chapter_patch.py; do not make one file-edit call per paragraph:

python ${HERMES_SKILL_DIR}/lib/validation.py chapter "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG" --chapter N

d. Mechanical slop scan:

!`python3 ${HERMES_SKILL_DIR}/lib/slop_detect.py "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/chapters/ch_{N:02d}.md" --genre-context "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/genre_context/resolved.json"`

Also run the manuscript-wide repetition scan (cheap, pure Python). Pass world.md, characters.md, canon.md, and the story contract as canonical references; never pass voice.md:

!`python3 ${HERMES_SKILL_DIR}/lib/slop_ngrams.py "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/chapters" --canonical-reference "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/world.md" --canonical-reference "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/characters.md" --canonical-reference "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/canon.md" --canonical-reference "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/genre_context/story-contract.md"`

Get the full breakdown (not just the score) with the same resolved patterns — the surgical rewrite loop needs the matched spans, not the number:

from genre_patterns import load_patterns_from_resolved
from slop_detect import find_spans, slop_report

genre_patterns = load_patterns_from_resolved(
    "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/genre_context/resolved.json"
)
spans = find_spans(chapter_text, genre_patterns=genre_patterns)
report = slop_report(chapter_text, genre_patterns=genre_patterns)

Use genre_context/compiled/evaluation-chapter.md for the accompanying profile-compliance and semantic genre-slop review. For selected genre patterns, an error finding blocks Gate 1 only when its total hits exceed max_count. Always report warning findings, but never let them block Gate 1.

e. Enforce (five gates — see references/surgical-rewrite.md):

Acceptance requires ALL FIVE gates:

1. Mechanical scan clean — slop_detect findings within the zero-tolerance/bounded limits classified in references/surgical-rewrite.md (Acceptance section).
2. Manuscript repetition clean for THIS chapter — no blocking_findings entry from slop_ngrams has an occurrence in the current chapter. Keep warning and domain findings in the report, but do not rewrite them.
3. Sentence-grading gate — grading_gate() passes (references/sentence-grading.md).
4. Profile compliance — no excluded-tag subject matter, no scene exceeding profile.rating (references/quality-rubric.md's Profile compliance section).
5. Reveal-budget compliance — list every new story fact first disclosed in the chapter and cite the exact May reveal clause or scene-card Reveal cell that authorizes it. Any uncited fact or any Must withhold disclosure blocks acceptance.

Flow:

• Run all five gates for the initial audit. If any gate fails, collect every current blocker into the bounded surgical process: two batch rewrite rounds for the chapter, then one deterministic cut/merge pass. Never multiply the budget by the number of spans. warning and domain repetition findings do not enter the rewrite batch.

• All five gates pass → compute the 1-10 trend score ONCE on the final text, run fact extraction, then let the validator record the score and progress in one atomic checkpoint:

  python ${HERMES_SKILL_DIR}/lib/validation.py checkpoint "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG" --chapter N --status accepted --score <score>
  

  Commit only after the command returns checkpointed: true.

• Chapter cannot converge after the bounded cleanup → preserve the draft and checkpoint it explicitly without a passing score:

  python ${HERMES_SKILL_DIR}/lib/validation.py checkpoint "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG" --chapter N --status needs_attention --note "<remaining blocking findings>"
  

  Continue only after checkpointed: true. Never write progress fields or the attention queue directly.

Step 3 — When all chapters drafted

1. Concatenate the chapter files in numeric order to a temporary manuscript file, then run the contextual scan to compute the novel-level score:

!python3 ${HERMES_SKILL_DIR}/lib/slop_detect.py "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/manuscript-draft.md" --genre-context "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/genre_context/resolved.json"

Remove the temporary file after recording the score.
2. Record via `record_score(s, "novel", <score>)`.
3. Update state.json `phase` to `"revision"`.
4. Commit. Hand off to `autonovel-revision`.

## Handoff

Invoke `autonovel-revision` when all chapters drafted.

## Library utilities

- `lib/slop_detect.py` — mechanical slop detection and full breakdown via `slop_report` (no LLM)
- `lib/slop_ngrams.py` — manuscript-wide repeated-phrase scan (no LLM)
- `lib/grading.py` — sentence-grading parse + deterministic acceptance gate
- `lib/state.py` — state.json management
- `lib/story_state.py` — story-state ledger: entities, facts with validity
intervals, foreshadow tracking, STORY STATE query block (no LLM; mirrored
into this skill's `lib/` in Module A3 — canonical source is
`autonovel/lib/story_state.py`)

## See also

- `references/chapter-prompt.md` — the chapter draft template
- `references/surgical-rewrite.md` — the enforcement loop and the five acceptance gates
- `references/sentence-grading.md` — the sentence-grading pass
- `references/retry-policy.md` — how to checkpoint a chapter after bounded cleanup
- `references/quality-rubric.md` — how to interpret the 1-10 trend score
- `references/fact-extraction.md` — post-acceptance fact extraction into the story-state ledger