Skip to content

Repository files navigation

Scribe

A Claude Code plugin that catches AI-default writing patterns before they ship.

Scribe routes every draft through a persona system, picks a varied subset of anti-patterns per run (so two runs never produce the same critique), scores prose on a 0-100 sameness scale, and fixes voice, rhythm, vocabulary, structure, and rhetoric before the book goes to print.

Built for longform authors, essayists, ghostwriters, and editors. Runs a team of specialist agents that plan, draft, edit, and review collaboratively.

The Problem

AI-generated writing converges on the same cadence. "Delve into the tapestry of." "It's important to note that." "Not just X, but Y." "This is not X. It is Y." Em-dashes everywhere. Three-word lists. Conclusion paragraphs that restate the previous five. Hedging openers.

Scribe fixes this with persona-first generation and a sameness score that quantifies how close prose is to the AI default.

Installation

Add the raymondware marketplace, then install Scribe:

/plugin marketplace add raymondware/kern
/plugin install scribe@raymondware

Restart Claude Code and the /scribe:* commands are available.

Quick Start

# Generate a passage from scratch
/scribe:write opening chapter for a popular nonfiction book about how cities forget their own histories

# Audit existing prose
/scribe:audit
# (paste the manuscript section to audit)

# De-AI an existing draft
/scribe:differentiate

# Run only the anti-pattern selector (verifies rotation)
/scribe:draw memoir essay about my year in Buenos Aires

Commands

Command Does
/scribe:write <desc> Full pipeline: persona detection, specialists, draft, review
/scribe:audit Anti-pattern scan with sameness score against existing prose
/scribe:revise Rewrite a passage against Scribe's standards, show diff
/scribe:differentiate Identify AI writing tells and apply targeted fixes
/scribe:diagnose Single-pass AI-ism detection, fast turnaround
/scribe:compare <a> <b> Side-by-side prose comparison with similarity score
/scribe:plan <desc> Writing specs only (voice, rhythm, structure) -- no draft
/scribe:review Full review pass: critic ensemble, voice editor
/scribe:research Scan writing communities for new anti-patterns
/scribe:draw <desc> Run the anti-pattern selector standalone (rotation verification)

How It Works

1. Persona Detection

Scribe picks a persona from your description, then loads that persona's rules for voice, rhythm, vocabulary, and structure. Five personas:

Persona For
literary-fiction Novels, novellas, short stories. Voice and rhythm primary.
popular-nonfiction Books that explain ideas to a general audience. Gladwell/Lewis/Pollan register.
technical-writing Books, docs, essays for technical readers. Precision over style.
memoir First-person life writing. Voice is everything.
business-book High AI-default risk. Hardest to make distinctive.

2. Variation Rule

Every run picks a different subset of anti-patterns. Scribe reads the audit log (state/draws.jsonl) and excludes the previous run's subset. Two runs on the same prompt produce different critiques.

3. Parallel Critics + Synthesizer

Five critics run in parallel (rhythm, voice, cliche, AI-ism, pacing). The synthesizer merges their output, removes duplicates, and computes a consensus sameness score.

4. The Sameness Score

Every draft gets a 0-100 score on AI-default similarity:

Score Meaning
0-20 Distinctive. A voice. Ship.
21-40 Has voice. Some defaults showing.
41-60 Generic. Could be any AI-generated draft.
61-80 Default AI prose. Multiple tells per page.
81-100 Unedited AI output. Do not ship.

Anti-Pattern Library

Scribe ships with anti-patterns across three files:

File Patterns
anti-patterns/language.md Delve/tapestry/realm vocabulary, tricolon reflex, hedge stacking, marketing verbs, em-dash overuse, self-descriptors, negative/positive restatement, filler intensifiers
anti-patterns/rhetoric.md Universal aphorisms, "not just X but Y" frames, throat-clearing openers, flashlight-hedging adjectives
anti-patterns/structure.md Conclusion-restating, predictable section sequence, thesis-delay syndrome

Agent Team

Specialists (Haiku)

rhythm-specialist voice-specialist vocabulary-specialist structure-specialist dialogue-specialist

Drafting (Sonnet)

draft-implementer voice-implementer

Critics (parallel ensemble)

rhythm-critic voice-critic cliche-detector ai-ism-critic pacing-critic

Orchestration (Opus 4.8)

conductor anti-pattern-selector critique-synthesizer

Contributing

See CLAUDE.md for the contributor guide.

License

MIT. See LICENSE.

About

Multi-agent writing harness that catches AI-isms.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages