Velith

한국어 · 日本語 · 中文 · Español · Français · Deutsch · Português

Books to the human-quality bar. A six-phase pipeline that takes a book from a blank page to a publishable EPUB and PDF, and holds every chapter, every edit, and every image to one standard: a cold reader who buys books in this genre cannot tell it was machine-drafted.

Phase 0: Onboarding → Phase 1: Ideation → Phase 2: Outlining → Phase 3: Drafting → Phase 4: Editing → Phase 5: Publishing

Why Velith?

Frontier models write good sentences. Left alone, they still write books that readers put down: a voice that drifts by chapter eight, characters who explain their feelings, statistics that do not exist, paragraphs that all end on a punch line, illustrations that change style every chapter. None of that is a model problem. It is a pipeline problem.

Velith is the pipeline. It reads the whole manuscript before writing the next chapter, locks the voice on a sample before drafting volume, critiques and revises every chapter before saving it, fact-checks every claim, rewrites during editing instead of filing reports, cold-reads the finished book as its target readers, and refuses to publish until they would keep reading. Images get the same treatment: one art bible per book, prompts compiled from it for whichever image model you use, diagrams rendered from code, and every picture opened and judged before it ships.

What changed in 0.5

Before (0.4)	Now (0.5)
Agents received chapter summaries	Agents read the entire manuscript (frontier models carry 1M tokens; a novel is under 200K)
Four chapters drafted in parallel	Narrative genres draft in order; chapter N reads chapter N-1
One pass per chapter	Draft → cold critique with quoted lines → revise, before the file is saved
Editing produced reports	Editing rewrites the manuscript in place, with snapshots, across 7 stages
"delve"-list slop detection	The 2026 AI-tell taxonomy: rhythm, structure, emotion, dialogue, plus en/ko/ja lexical tells, measured by velith.mjs metrics
No fact-checking	fact-checker builds a claim ledger and removes what it cannot verify
Gate = files exist	Gate = beta-reader readiness verdict: five axes ≥ 7, no put-down point in chapters 1-3
Cover prompts only	Art bible, look lock, code-rendered figures, compiled prompts for any backend, vision QA, asset validation

	Feature	Why it matters
📏	One quality bar	skills/loom/quality-bar.md: five-axis rubric, AI-tell taxonomy, cold-read protocol. Every agent reads it
📋	6-phase pipeline with author checkpoints	Concept, outline, voice lock, look lock, restructures, readiness. Everything else runs unattended
📖	7 genre craft references + custom	Fiction, non-fiction, technical, screenplay, poetry, game, academic: structure options, craft, genre-specific tells, language notes
🤖	12 specialized agents	Architect, scene planner, writer, continuity, fact-checker, style doctor, beta reader, art director, figure engineer, illustrator, cover, marketing
✏️	7-stage editing	Fact check → assessment → developmental → line → copy → proofread → readiness
🎨	Visual system	Art bible, look lock, Mermaid/D2/SVG figures, model-agnostic prompts, vision QA, print/EPUB checks
📊	Deterministic metrics	Sentence rhythm, paragraph shape, repetition across chapters, tell density (en/ko)
📦	EPUB, PDF, MOBI, TXT, Markdown	Pandoc + optional Calibre, epubcheck, KDP and Korean platform checklists

Comparison

	Velith	Raw prompts	Notion AI	Jasper / Sudowrite	Scrivener
Full-manuscript context	Every agent, every task	Manual	None	Limited	n/a
Voice lock + critique-revise loop	Built in	None	None	None	Manual
Fact-checking with claim ledger	Dedicated agent	None	None	None	Manual
Readiness gate by simulated readers	Blocks publishing	None	None	None	None
Visual consistency across a book	Art bible + compiled prompts + vision QA	Per-image prompts	None	None	None
Genre awareness	7 craft references + custom	Prompt-dependent	None	Fiction-focused	None
Output format	EPUB, PDF, MOBI, TXT, Markdown	Copy-paste	Markdown / PDF	DOCX, limited	DOCX, PDF
Requires	Claude Code, Codex CLI, Grok Build, Agy, Cursor, Cline, or Aider	Any LLM	Notion subscription	Subscription	License
Full control	Prompt-level, Apache-2.0	Full	Black box	Black box	Full

Installation

Claude Code

/plugin marketplace add epicsagas/plugins
/plugin install velith@epicsagas

All 18 skills and 12 agents are available immediately. Updates with /plugin update velith@epicsagas.

Prerequisites: Claude Code CLI installed and authenticated. Velith is tuned for the Claude 5 family (1M context); agents inherit your session model and set their own effort level.

Codex CLI (OpenAI)

codex plugin marketplace add epicsagas/plugins

Velith provides 18 skills and 12 custom subagents (.codex-plugin/agents/*.toml, generated from agents/*.md). Codex auto-discovers both. Updates with codex plugin update velith@epicsagas.

Prerequisites: Codex CLI installed and configured.

Grok Build (xAI)

grok plugin install epicsagas/Velith --trust

Grok reads skills from skills/ and agents from agents/ at the plugin root. No extra configuration needed. Agents spawn as velith:<name> (for example velith:chapter-writer).

Alternatively, add this repository as a marketplace:

grok plugin marketplace add epicsagas/Velith
grok plugin install velith --trust

Updates with grok plugin update velith.

Prerequisites: Grok Build installed and authenticated.

Agy (Antigravity)

agy plugin install https://github.com/epicsagas/Velith

Cursor

Context rules in .cursor/rules/:

Rule File	Loaded When
velith-pipeline.mdc	Always (phases, router, agents, quality bar, checkpoints)
velith-genres.mdc	Editing drafts, outlines, or PRD files
velith-editing.mdc	Working on edits, STYLE.md, or bible.md

Cline

Project-level instructions in .clinerules at the repository root.

Aider

Writing conventions in CONVENTIONS.md, auto-loaded via .aider.conf.yml.

Quick Start

> /book-init          # genre, reader, language, voice sample → PRD.md + STYLE.md
> /loom               # detects state and runs the next phase, stopping at author checkpoints

What happens:

1. Onboarding — reader, promise, scope, and a voice fingerprint from your own writing sample
2. Ideation — premise stress test, real comparable titles, ranked concepts; you choose
3. Outlining — structure chosen and justified, chapter specs, figure plan, bible; architect scores it; you approve
4. Drafting — voice lock on a sample chapter; then sequential full-context chapters, each critiqued and revised, bible ledger updated, continuity checked
5. Editing — fact check, assessment, developmental rewrites, line edit, copy edit, proofread, then a cold read by simulated target readers. PASS or REVISE
6. Publishing — front and back matter, EPUB/PDF/MOBI, epubcheck, cover from the art bible, marketing plan, platform checklists

Images at any point from Phase 2: /book-visuals plan (art bible), /book-visuals lock (look lock), then figures and illustrations as the chapters need them.

Skills

Skill	Phase	Description
/loom	Router	Detect state, run the next phase, enforce gates
/book-init	0	Reader, promise, scope, voice fingerprint, source index → PRD.md, STYLE.md
/book-ideation	1	Premise stress test, comps, ranked concepts, voice samples
/book-outline	2	Structure, chapter specs, figure plan, bible, scored validation, approval
/book-draft	3	Voice lock, sequential draft-critique-revise, ledger, continuity
/book-edit	4	7 stages: fact check … readiness verdict
/book-publish	5	Readiness gate, matter, formats, epubcheck, cover, marketing, checklists
/book-visuals	2-5	Art bible, look lock, figures, illustrations, photos, prompt compile, vision QA, asset check
/book-illustrate	3-5	Alias for the illustration subset of /book-visuals
/book-status	—	Terminal dashboard, --ui browser dashboard, --metrics
/book-fiction … /book-academic	—	Genre craft references (7)
/book-genre-creator	—	Genre selection and custom genre specs

Agents

Agent	Phase	Job
book-architect	2	Outline and bible; structure chosen and justified; scored validation; restructure proposals
scene-generator	3	Scene plans per fiction chapter (purpose, turn, subtext, exit). Plans, not prose
chapter-writer	3-4	One chapter in full context; drafts, critiques with quoted lines, revises, updates the ledger
continuity-editor	3-4	Contradictions and repetitions across the whole manuscript against the bible
fact-checker	4	Claim ledger; verification against sources and web; removes what cannot be verified; runs code
style-doctor	4	Measures, then rewrites tells, rhythm uniformity, and drift in place
beta-reader	4	Cold read as three target readers and one professional; readiness verdict
art-director	2-5	Art bible, look lock, vision QA of every image, contact-sheet review
figure-engineer	3-5	Diagrams, charts, technical drawings, maps' base layers as code; labels verified against the text
illustrator	3-5	Illustrations from the art bible with compiled prompts, generation when a tool exists, vision QA
cover-designer	5	Cover concepts from the art bible, formats, marketing variants
marketing-expert	5	Positioning, personas, channels, calendar, launch checklist

The quality bar

skills/loom/quality-bar.md is read by every writing, editing, and reviewing agent. It defines:

• Five axes, scored 1-10 with anchors: voice and prose, structure and pacing, depth, specificity and grounding, reader experience. Voice lock needs axis 1 ≥ 7. Readiness needs every axis ≥ 7, mean ≥ 7.5, and no put-down point in the first three chapters.
• The 2026 AI-tell taxonomy: what readers actually notice now (uniform paragraph rhythm, punch-line closures, "not X but Y", reflective codas, precise self-knowledge, dialogue that answers, hedged authority, invented specificity) with fixes, plus lexical lists for English, Korean, and Japanese.
• The cold-read protocol: read once at speed, mark, diagnose with quoted lines, score honestly, prioritize, then revise.
• Visual rules: one look per book, purpose before decoration, code for anything with text or data, look lock, vision QA, release constraints.

CLI

node velith.mjs scan <dir> [--ui]           # project state, dashboard data, readiness verdict
node velith.mjs metrics <dir|file>          # prose metrics + cross-chapter repetition (JSON)
node velith.mjs snapshot <dir> <label>      # copy drafts/ before a rewriting stage
node velith.mjs images compile <dir> [id]   # art bible + spec → Midjourney / gpt-image / SD-FLUX / Imagen / Ideogram prompts
node velith.mjs images check <dir>          # dimensions, aspect, size, alt text, references, manifest coverage
node velith.mjs images render <dir>         # Mermaid / D2 / Graphviz / SVG / matplotlib → SVG + PNG

Visual Dashboard

/book-status --ui opens a Svelte dashboard: pipeline tracker, 12 agent cards, chapter table, 6-stage editing kanban, readiness verdict with axis scores, output files, settings. Pre-built dist/ is included.

cd dashboard && npm install && npm run dev   # http://localhost:5173
npm run build                                 # rebuild dist/

External Dependencies

brew install pandoc                 # EPUB/PDF (required for Phase 5)
brew install texlive                # PDF with CJK support
brew install --cask calibre         # MOBI (optional)
brew install epubcheck              # EPUB validation (optional, recommended)
npm i -g @mermaid-js/mermaid-cli    # Mermaid figures → SVG (optional)
brew install d2 graphviz librsvg    # D2 / Graphviz figures, SVG → PNG (optional)

Image generation is not bundled. illustrator and cover-designer generate when an image tool is available in your session (MCP image servers, Replicate, local Stable Diffusion); otherwise they hand you compiled prompt packs per backend in visuals/prompts/.

Troubleshooting

• pandoc not found — brew install pandoc
• CJK characters missing in PDF — brew install texlive or brew install basictex && sudo tlmgr install collection-langkorean
• Plugin commands not found — restart Claude Code
• Phase 4 never completes — the gate is edits/readiness-report.md with verdict: PASS; run /book-edit 6
• Images look different per chapter — no look lock; run /book-visuals plan then /book-visuals lock

Project Structure

{project-dir}/
├── PRD.md              # Requirements + reader promise
├── STYLE.md            # Voice fingerprint, voice sample, rules, voice lock
├── ideation.md         # Concepts, comps, chosen concept
├── outline.md          # Chapter specs, figure plan, validation, approval
├── bible.md            # Characters/concepts, term rules, timeline, per-chapter ledger
├── art-bible.md        # Visual identity, figure system, look lock
├── sources/            # Reference material + INDEX.md
├── drafts/             # ch{NN}-{slug}.md, ch{NN}-scenes.md (revised in place during editing)
├── visuals/            # plan, manifest, figures/, illustrations/, photos/, ref/, prompts/
├── edits/              # 00-fact-check … 06-readiness-report, readiness-report.md, editorial-report.md
├── publish/            # book.epub/pdf/…, metadata, front/back matter, cover/, marketing, checklists
└── .velith/            # status.json, art-bible.json, critiques/, snapshots/, metrics.json

Integration

• alcove — search your document vault as source material during /book-init and drafting.
• obsidian-forge — of book init / sync / export to write from an Obsidian vault.
• humanize-korean — if installed, style-doctor may run it as a final Korean polish.
• Image generation MCPs — used by illustrator and cover-designer when present.

All optional. Velith works standalone.

Contributing

See CONTRIBUTING.md. Prompts are the product: a change to quality-bar.md or an agent file changes every book. Test with examples/ and node velith.mjs metrics.

License

Apache-2.0