autonovel
A first-party Nous Research companion repo: an end-to-end autonomous novel-writing pipeline orchestrated by Hermes Agent, turning a seed concept into a print-ready PDF, ePub, audiobook, and landing page. It is the most concrete deployment evidence in the Hermes orbit — a working pipeline with a shipped artifact — and exactly the kind of first-party primary source the awesome-hermes-usecases catalog requires (content-creation category).
The pipeline — four phases, ~27 scripts
- Foundation — world bible, character registry, outline, voice fingerprint, canon; loops until quality > 7.5.
- First draft — chapters written sequentially, kept if score > 6.0 else retried; forward progress over perfection.
- Revision — (a) automated: adversarial edit → cut → reader-panel eval → brief → rewrite, with plateau detection as the stop condition; (b) an Opus review loop — the full manuscript goes to Claude Opus under a dual-persona prompt (critic + fiction professor) until major issues resolve.
- Export — LaTeX typesetting, cover art, audiobook, landing page.
The verification core — two “immune systems”
What makes it a loop rather than a one-shot generator is its feedback signal:
- Mechanical scan — regex rules for banned words, clichés, show-don’t-tell violations.
- LLM Judge — a separate model scoring prose quality and voice adherence. This is the spoke’s “the bottleneck is verification, not generation” thesis made literal: goal + stop-condition (>7.5 / >6.0 / plateau) + a truthful-enough feedback signal, iterated.
Model-agnostic in practice
Though “by Hermes Agent,” the content models are Anthropic Claude (Sonnet to draft/revise, Opus to review), fal.ai (Nano Banana 2 cover art), and ElevenLabs (multi-voice narration) — a concrete demonstration of hermes-agent‘s model-agnostic orchestration: Hermes is the conductor, Claude the pen.
Production evidence (the artifact exists)
First novel — The Second Son of the House of Bells, published under the byline “Claude Hermes” (reader-facing landing page, PDF/ePub/audiobook): 79,456 words / 19 chapters (merged from 24), 6 automated revision cycles + 6 Opus review rounds, 19 woodcut ornaments, and a full audiobook of 4,179 speaker-attributed segments. The repo ships reusable craft docs (CRAFT.md, ANTI-SLOP.md, PIPELINE.md) and credits Karpathy’s autoresearch iteration paradigm, applied to fiction — where autoresearch had an objective metric (val_bpb), autonovel had to build its feedback signal.
Tier
T1 — first-party project repo (the code + a reproducible shipped artifact), the spoke’s convention
for project source. Self-interested (Nous promoting Hermes), but the output novel is concrete, not a
marketing claim; weakness recorded. freshness: volatile (active repo).
Related
second-son-house-of-bells · hermes-agent · nous-research · awesome-hermes-usecases · autoresearch · agent-loops-verification · loop-engineering · agent-orchestration · self-improving-agents · claude-opus-4-8