Autonomous fantasy-novel pipeline built as a Claude Code project — skills, subagents, and slash commands instead of a Python script. Same modify → evaluate → keep/discard loop, applied to fiction; same five-layer co-evolving structure (voice / world / characters / outline / chapters, plus canon).
Scope. Phases 1–3: foundation → drafting → automated revision → Opus review. Art, audiobook, ePub, landing page, and LaTeX typesetting are out of scope.
⚠️ Token cost. A full run from seed to Phase 3b finish takes several hours and burns serious tokens (Opus review is the heaviest step). Inspectexamples/ashen-crown/first to see what intermediate artifacts look like before committing to a run.
git clone https://github.com/tchr-dev/autonovel.git
cd autonovelclaudeSuppresses prompts during the multi-hour run:
cp .claude/settings.example.json .claude/settings.local.jsonSkills under .claude/skills/ and slash commands under .claude/commands/
are auto-discovered.
No API key is required for the pipeline itself — Claude Code provides the
LLM. .env.example is preserved only for users who want to run the
standalone Python helpers in scripts/ against the Anthropic API
directly.
In Claude Code:
/autonovel my-novel-tagThis will:
- Create
novels/my-novel-tag/and initialise state. - Ask for a seed concept (or invoke the
seedskill to generate options). The rootseed.txtis a drop-in template you can edit. - Run Phase 1 (Foundation) until
foundation_score > 7.5 AND lore_score > 7.0. - Run Phase 2 (Drafting) chapter by chapter until all are above 6.0.
- Run Phase 3a (Automated revision cycles) until scores plateau.
- Run Phase 3b (Opus review loop) until no major unqualified items remain.
State persists in novels/<tag>/state.json. Resume with
/autonovel-resume <tag>.
examples/ashen-crown/ is a real partial run —
foundation complete, one chapter drafted, evaluation logs intact. Browse
it to see what each phase produces, or copy it into novels/ and resume:
cp -r examples/ashen-crown novels/ashen-crown
# Then in Claude Code:
/autonovel-resume ashen-crownThe seed that produced this run lives at both
seed.txt (repo root, as a template) and
examples/ashen-crown/seed.txt.
The output of this pipeline is built specifically to dodge the structural and lexical tells that fire AI detectors. The first drafted chapter of the worked example was run through two public detectors:
| Detector | Sample | Result | Screenshot |
|---|---|---|---|
| Yandex Neurodetector | 2,154 words / 11,133 chars / 10 segments | 2.86% AI-generated probability |  |
| GPTZero (Model 4.4b) | 2,154 words / first 10,000 chars scanned | AI 0% · Mixed 0% · Human 100% — "highly confident this text is entirely human" |  |
The framework references in framework/ANTI-SLOP.md
and framework/ANTI-PATTERNS.md, combined
with the regex slop_scan.py pass and the reader-panel evaluation, are
what drive these numbers down. Detector scores are not a quality
guarantee, but they're a useful sanity check that the prose has avoided
the obvious AI fingerprints.
.claude/
skills/ # the LLM-driven steps Claude executes
autonovel/ # top-level orchestrator
seed/
foundation/ # gen-world, gen-characters, gen-outline,
# gen-canon, voice-discovery
drafting/ # draft-chapter
evaluation/ # evaluate-{foundation,chapter,full},
# adversarial-edit, compare-chapters,
# voice-fingerprint
revision/ # gen-brief, gen-revision, reader-panel,
# opus-review, apply-cuts
agents/ # personas spawned as subagents
panel-{editor,genre-reader,writer,first-reader}.md
literary-critic.md
professor-of-fiction.md
commands/ # slash commands
autonovel.md, autonovel-resume.md
settings.example.json # pre-approved permissions for unattended runs
scripts/ # mechanical (no LLM)
slop_scan.py # regex slop scanner
apply_cuts.py # quote-match cut applicator
voice_fingerprint.py # quantitative voice analysis
build_manuscript.py # concatenate chapters
parse_review.py # parse Opus review markdown
state.py # state.json read/update
results_log.py # append to results.tsv
framework/ # immutable craft references
CRAFT.md
ANTI-SLOP.md # word-level AI tell detection
ANTI-PATTERNS.md # structural AI pattern detection
voice-guardrails.md # universal voice rules (Part 1)
templates/ # copied per novel
voice.md # full skeleton (Part 1 + blank Part 2)
world.md, characters.md, outline.md, canon.md, MYSTERY.md
state.json
examples/ # committed worked runs (read-only references)
ashen-crown/ # foundation done + 1 chapter drafted
novels/<tag>/ # one per novel run (gitignored)
seed.txt
voice.md, world.md, characters.md, outline.md, canon.md, MYSTERY.md
chapters/ch_01.md, ch_02.md, …
arc_summary.md, manuscript.md
state.json, results.tsv
edit_logs/ # adversarial cuts, panel JSON, tournament
eval_logs/ # evaluate-* outputs
reviews/ # Opus review markdown + parsed JSON
briefs/ # revision briefs
.snapshots/ # phase-level rollback points
- Skills replace LLM-calling Python scripts. No httpx boilerplate, no JSON parsing, no model-name configuration. Claude executes the writing and evaluation directly.
- Subagents replace persona-prompted httpx calls. Each reader-panel persona and each Opus reviewer runs as an isolated subagent context — independent reasoning, no cross-contamination.
novels/<tag>/directories replace per-novel git branches. Snapshots are explicit, taken at phase transitions and before risky rewrites. Nogit reset --hardfor "discard"; restore from.snapshots/instead.- State is JSON, not git history. Resume from
state.json.
- The five-layer co-evolving structure (voice / world / characters / outline / chapters + canon)
- Two immune systems: mechanical regex (
slop_scan.py) + LLM judge (evaluate-*skills) - Reader panel of four personas, with consensus / disagreement detection
- Opus dual-persona review loop with stopping conditions
- Adversarial "cut 500 words" pass
- Swiss-tournament Elo ranking of chapters
# Foundation
gen-world, gen-characters, gen-outline, gen-canon, voice-discovery
evaluate-foundation
# Drafting
draft-chapter (per chapter)
evaluate-chapter
# Revision
adversarial-edit (per chapter or all)
apply-cuts (mechanical)
compare-chapters (tournament)
voice-fingerprint (mechanical)
reader-panel (4 subagents)
gen-brief, gen-revision
opus-review (2 subagents)Skills assume an Opus-class model is available for Phase 3b review and a Sonnet-class model for drafting/evaluation. On a Haiku-only plan expect degraded prose and weaker review signal.
Subagent model selection is pinned in agent frontmatter rather than
inherited from the orchestrator session, so the right model runs on each
persona regardless of which model you launched /autonovel from:
| Agent | Pinned model | Rationale |
|---|---|---|
literary-critic, professor-of-fiction |
opus |
Full-manuscript read; long-form criticism and severity-tagged punch list. The professor's severity tags drive the Phase 3b stopping condition, so review depth has systemic impact. |
panel-editor, panel-writer, panel-genre-reader, panel-first-reader |
sonnet |
Read an arc summary (not the full manuscript) and emit structured JSON. Sonnet handles this well without burning Opus budget. |
For very long manuscripts (≳120k words) where the standard 200k Opus
window crowds, switch the two Opus agents to claude-opus-4-7[1m]. This
is opt-in per run — the default stays at standard Opus to avoid paying
1M-context rates on the heaviest phase of every pipeline.
MIT — see LICENSE.