Ghostwriter

Ghostwriter — Claude Code Plugin

An adversarial AI text detection and humanization system. Detects AI-generated text and iteratively revises it to read as authentically human.

How It Works

The ghostwriter uses an adversarial loop: a Writer agent produces text, a Detector evaluates it for AI signals, and an AI Engineer refines the Writer's approach based on feedback. Each round, the text gets harder to distinguish from human writing.

Writer → Reviewer → Detector → Pass? → Done
                                  ↓
                            AI Engineer → next round

Requirements

• bun (required) — Install bun
• Python 3.8+ (optional) — enables additional detection tools (burstiness, n-gram, syntactic complexity, content analysis, perplexity). Without Python, only TypeScript-based tools run.

Installation

Local development

claude --plugin-dir /path/to/ghostwriter

Quick Start

1. Setup

/setup

The setup wizard guides you through creating your author profile, naming your publication, and choosing your writing medium and style.

2. Analyze text

Check if text reads as AI-generated:

/analyze guides/getting-started.md developer-docs
/analyze path/to/text.md

3. Quick rewrite (single pass)

Fast humanization without the iterative detection loop:

/humanize guides/getting-started.md developer-docs --quick
/humanize input.md output.md --quick

4. Humanize text (iterative loop)

Full adversarial loop with detection, review, and improvement:

/humanize guides/getting-started.md developer-docs
/humanize input.md output.md <publication-name>

Available Commands

Command	Description
/setup	Interactive setup wizard — create authors and publications
/audit <publication>	Scan all files in a publication for AI signals
/analyze <file> [publication]	Analyze a single file for AI signals
/humanize <file> <publication>	Iterative humanization loop (Writer → Reviewer → Detector → AI Engineer)
/humanize <file> <publication> --quick	Single-pass rewrite without detection loop
/humanize-all <publication>	Batch humanize all files in a publication
/create <topic> [--config PATH]	Generate new text from a topic via adversarial training
/authors [list|add|remove|show]	Manage author profiles
/publications [list|add|remove|show]	Manage publications
/migrate	Migrate from v1 layout to v2
/help	Show available commands and quick start guide

Project Structure

After setup, your project will look like:

my-project/
├── .ghostwriter/
│   ├── learned-patterns/        # Accumulated training patterns
│   │   ├── global.md
│   │   ├── blog.md
│   │   └── book.md
│   ├── authors/
│   │   └── your-name/
│   │       ├── author-persona.yml   # Your author profile
│   │       └── learned-patterns.md  # Author-specific patterns
│   └── publications/
│       └── my-publication/
│           ├── config.yml           # Publication config
│           ├── learned-patterns.md  # Publication-specific patterns
│           └── pipeline/            # Humanization run data
├── apps/docs/src/content/    # Your content (untouched)
└── content/blog/             # Your content (untouched)

An "author" represents a writing persona — it can be a real person, a team ("Acme Engineering"), or a brand voice.

Configuration

Publication configs use YAML with a preset system:

version: '1.0'

presets:
  - ../../authors/your-name/author-persona.yml
  - technical-book

publication:
  name: "My Technical Book"
  media: book
  audience: expert

content:
  root: ../../../manuscript
  glob: "**/*.md"

writing_style:
  tone: conversational
  formality: informal
  voice: first-person

Available Presets

blog, book, technical-book, academic, business, casual, technical-docs, student

Templates

The plugin ships templates for common writing scenarios. Run /setup to use them interactively, or copy from the plugin's templates/ directory.

Learned Patterns

The system learns from each humanization round. Patterns are stored in a layered hierarchy:

1. Global — .ghostwriter/learned-patterns/global.md — applies to all writing
2. Medium — .ghostwriter/learned-patterns/{media}.md — applies to a specific medium
3. Medium+Style — .ghostwriter/learned-patterns/{media}-{style}.md — applies to a specific style
4. Author — .ghostwriter/authors/{slug}/learned-patterns.md — applies to all writing by this author
5. Publication — .ghostwriter/publications/{slug}/learned-patterns.md — applies to this publication only

Later layers override earlier ones on conflict.

License

MIT