Spec-driven writing, publishing, and translation for AI coding agents.
I don't outline -- Claude does. I don't edit -- Claude does. I don't format -- Claude does. I just write.
Scriven brings spec-driven workflows to longform writing. It runs inside your AI coding agent (Claude Code, Cursor, Gemini CLI, and more) and also ships a guided Perplexity Desktop setup path for file-aware support without overstating parity.
Scriven is best understood as AI-native longform writing software built around voice preservation. Its core promise is narrow and high-stakes: drafted prose should sound like the writer, not like AI. If you want evidence before features, start with the Proof Artifacts.
npx scriven-cli@latestScriven is a command system that turns your AI coding agent into a voice-preserving writing studio. It supports 50 work types -- novels, screenplays, research papers, technical guides, runbooks, scripture commentaries, comics, memoirs -- each with its own adaptive vocabulary and toolset.
The wedge comes first: Scriven profiles the writer, loads that voice into every drafting step, and keeps each unit on fresh context so the prose stays specific to the project. From there, it expands into 108 writing commands covering the rest of the pipeline:
- Create -- Set up a project with tailored context files. Progressive onboarding, never overwhelming.
- Write -- Discuss, plan, draft, and revise one unit at a time. The drafter agent loads your Voice DNA and writes in your voice, not generic AI prose.
- Polish -- Editor review, line edit, copy edit, continuity check, voice check, beta reader simulation, sensitivity review.
- Publish -- Front/back matter, cover art, blurbs, query letters, KDP packages, IngramSpark packages, EPUB, PDF, Fountain, Final Draft, LaTeX.
- Translate -- Deep translation with glossary management, cultural adaptation, back-translation verification, multi-language simultaneous publishing.
- Collaborate -- Parallel revision tracks, co-writing workflows, continuity merge checking.
Everything adapts to your work type. A novel uses /scr:draft for chapters. A screenplay uses /scr:draft for acts. A Quran commentary uses /scr:draft for surahs. Same command, tradition-native vocabulary.
# Install
npx scriven-cli@latest
# In Claude Code:
/scr-new-work # Start a fresh project
/scr-demo # Explore a pre-built sample first
/scr-next # The universal "what should I do now" command
/scr-help # See what's available for your work type
# Other slash-command runtimes currently keep /scr:*:
/scr:new-work
/scr:next
# In Codex, use the generated $scr-* skills:
$scr-new-work
$scr-demo
$scr-next
$scr-helpIf you only ever type /scr-next in Claude Code, you can complete an entire novel. It always knows what's next.
If you want the shortest proof-first route, read Proof Artifacts before exploring the rest of the docs.
Scriven's core insight: drafted prose should sound like you, not like an AI. Before drafting begins, /scr:profile-writer builds a detailed voice profile across 15+ dimensions:
- Narrative perspective, tense, narrator stance
- Sentence architecture, paragraph rhythm
- Vocabulary register, figurative density, recurring image systems
- Dialogue style, character voice differentiation
- Pacing, transitions, emotional range
- Do/don't/consider rules specific to the writer
This profile is saved as STYLE-GUIDE.md and loaded into every drafter agent invocation. The drafter writes one atomic unit per fresh context -- a scene, a subsection, a passage -- with the style guide as its primary reference. Voice stays consistent across hundreds of scenes.
For sacred and historical texts, Voice DNA is supplemented by 10 sacred voice registers (prophetic, wisdom, legal, liturgical, narrative-historical, apocalyptic, epistolary, psalmic, parabolic, didactic).
Prose: novel, novella, short story, flash fiction, memoir, creative nonfiction, biography, essay, essay collection
Script: screenplay, stage play, TV pilot, TV series bible, audio drama, libretto/musical
Academic: research paper, thesis/dissertation, journal article, white paper, literature review, monograph
Technical writing: technical guide/user guide, runbook/SOP, API or CLI reference, design spec/architecture doc
Visual: comic, graphic novel, children's book, picture book
Poetry: poetry collection, single poem, song/lyric
Interactive: interactive fiction, game narrative
Sacred & historical: scripture (Biblical, Quranic, Torah, Vedic, Buddhist, generic), commentary/exegesis, devotional, liturgical text, historical chronicle, historical account, mythological collection, religious epic, sermon, homiletic collection
Each work type has its own structural hierarchy and industry-standard word count and page range guidance -- a novel targets 70,000–100,000 words across 20–35 chapters, a screenplay targets 90–120 pages across 3–5 acts. These ranges guide outlining, progress tracking, and drafter pacing. The runnable command ids stay stable, while Scriven adapts the wording around them -- a Torah commentary still runs /scr:plan 3, but frames that work as planning Parashah 3.
Run the full pipeline autonomously. Three profiles:
- Guided -- Pause after each unit for review
- Supervised -- Batch through several units, pause for review
- Full-auto -- Run until complete; only pause on critical failures
# Claude Code
/scr-autopilot --profile supervised
# Other slash-command runtimes
/scr:autopilot --profile supervisedPlus:
/scr:autopilot-publish-- One command generates front matter, back matter, cover, and full export package/scr:autopilot-translate french german spanish-- Simultaneous multi-language editions
Scriven detects non-technical writers and hides git terminology. Instead of commit, branch, merge, diff -- you see save, version, compare, accept changes. All the power, none of the coding jargon.
Technical writers can enable developer mode in settings for full git access and verbose output.
Scriven is built on five principles:
-
The writer's voice is sacred. The drafter never imposes generic AI style. Every drafted sentence passes through the Voice DNA gate.
-
Fresh context per atomic unit. Each scene, subsection, or passage is drafted in a clean context. This prevents voice drift, context bloat, and keeps each unit at its best.
-
Progressive disclosure. Onboarding asks 3 questions, not 30. Depth is optional and always additive.
-
Tradition-native vocabulary. A Quran commentary uses surahs and ayahs. A Bible study uses books and verses. A screenplay uses acts and scenes. The tool adapts to the tradition -- the writer never adapts to the tool.
-
/scr-nextalways works in Claude Code. The universal interface. A writer who only ever types/scr-nextcan complete an entire novel, from blank page to KDP package.
- Proof Artifacts -- Canonical proof hub for the watchmaker sample flow and Voice DNA before/after bundle
- Getting Started -- Install to first draft in 10 minutes
- Command Reference -- All 108 commands with usage, flags, and examples
- Work Types Guide -- How 50 work types adapt Scriven's vocabulary
- Voice DNA Guide -- The 15+ dimension voice profiling system
- Publishing Guide -- 13 export formats, KDP, IngramSpark, submission packages
- Sacred Text Guide -- 15 sacred work types, voice registers, tradition-native features
- Translation Guide -- Multi-language pipeline with glossary and cultural adaptation
- Contributing -- How to add commands, agents, work types, and templates
- Architecture -- How Scriven works under the hood
- Configuration -- Package, installer, constraints, and
.manuscript/config.jsonsurfaces - Development -- Contributor workflow for changing commands, templates, installer logic, and docs
- Testing -- What the test suite covers and which checks to run before shipping
- Release Notes -- Public summary of what changed between package releases
- Shipped Assets -- Canonical inventory of bundled export templates and launch-critical files
- Runtime Support -- Canonical runtime matrix, Node baseline, and verification-status framing
Scriven currently ships installer targets for these AI tooling environments:
- Claude Code (primary reference runtime)
- Cursor
- Gemini CLI
- Codex
- OpenCode
- GitHub Copilot
- Windsurf
- Antigravity
- Manus Desktop
- Perplexity Desktop (guided local-MCP setup)
- Generic (SKILL.md) fallback
Installer baseline: Node.js 20+ for npx scriven-cli@latest and bin/install.js.
Support note: Claude Code is the primary reference runtime and now installs a flat /scr-* command surface. The environments listed above are installer targets, not a claim that every host runtime has verified parity today. Codex currently installs a skill-native $scr-* surface, while Perplexity Desktop is a guided local-MCP target rather than a writable command runtime. See the runtime compatibility matrix for install type, support level, and verification status.
Version: 1.5.3
Scriven's core command surface is stable across 108 commands, 50 work types, and 11 installer targets. The current repo baseline includes shipped planning milestones through v2.0 Publishing Cover Packaging, and the published package version is now 1.5.3. See Shipped Assets for the canonical asset inventory and Runtime Support for the runtime compatibility matrix.
Version 1.5.3 packages the post-1.5.2 hardening work that landed in the repo baseline: command-surface truthfulness, workflow-contract integrity, the Claude /scr-* surface, the Codex $scr-* skill surface, save-history contracts, session boundaries, and help/docs alignment, all without changing the zero-dependency architecture.
Package history is tracked in CHANGELOG.md, and the public-facing summary for this release is in docs/release-notes.md.
MIT. See LICENSE.
Scriven is an open project. Contributions welcome -- especially new work types, additional runtime adapters, and voice register definitions for languages and traditions we haven't covered yet.