Ghostwriter

Describe what you want to publish. Ghostwriter handles the rest.

Content creators, solopreneurs, and small teams spend 5-10 hours per article researching, writing, editing, and publishing across platforms. Most AI writing tools produce generic slop that readers immediately recognize as machine-generated.

Ghostwriter is an autonomous content engine that generates high-quality, on-brand content from real data sources — and actually sounds like you.

ghostwriter create "a weekly tech blog about AI safety, published to my WordPress site"

One command. Scheduled articles. Multiple platforms. Your voice.

---

The Problem

Pain point	What people do today	What Ghostwriter does
Content takes forever	5-10 hrs/article: research, write, edit, format, publish	End-to-end pipeline runs in minutes
AI writing sounds like AI	"Delve into the ever-evolving landscape..."	4 review agents + 93-phrase blacklist reject slop before publish
Voice is inconsistent	Manually rewrite AI output to match your tone	Style fingerprinting extracts your voice from existing writing — no fine-tuning
Multi-platform is tedious	Copy-paste and reformat for each platform	One piece adapts to blog posts, Twitter threads, and podcast scripts
No feedback loop	Guess what works, repeat what doesn't	Analytics sync feeds engagement data back into generation

---

How It Works

flowchart LR
    subgraph Input
        A["🗣️ Natural language\ndescription"] --> B["📋 Channel Config"]
        S["✍️ Your existing\nwriting samples"] --> F["🔍 Style\nFingerprint"]
    end

    subgraph Ingest["Data Ingestion"]
        B --> D["RSS Feeds\nAPIs\nWeb Scraping"]
        D --> |"cached &\ndeduped"| E["Source\nMaterial"]
    end

    subgraph Pipeline["Content Pipeline (LLM)"]
        direction TB
        E --> R["Research\n(Sonnet)"]
        R --> O["Outline\n(Sonnet)"]
        O --> DR["Draft\n(Opus)"]
        F -.->|"voice\ninjected"| DR
        P -.->|"engagement\ninsights"| DR
        DR --> RV{"4 Review\nAgents"}
        RV -->|"pass"| AD["Adapt"]
        RV -->|"fail"| PO["Polish &\nRetry"]
        PO -->|"≤3 tries"| RV
    end

    subgraph Publish["Multi-Platform Publishing"]
        AD --> WP["WordPress"]
        AD --> TW["Twitter\nThreads"]
        AD --> POD["Podcast\n(TTS)"]
    end

    subgraph Feedback["Analytics Feedback Loop"]
        WP & TW & POD --> AN["Engagement\nMetrics"]
        AN --> P["Performance\nInsights"]
    end

    style Input fill:#f0f4ff,stroke:#4a6fa5
    style Ingest fill:#fff4e6,stroke:#c48a2a
    style Pipeline fill:#f0fff4,stroke:#3a8a5c
    style Publish fill:#fff0f0,stroke:#a54a4a
    style Feedback fill:#f5f0ff,stroke:#6a4aa5

Loading

The pipeline in plain English:

1. You describe a channel — topic, tone, audience, schedule, platforms
2. Ghostwriter ingests real data — RSS feeds, market APIs, recipe databases — cached and deduplicated
3. Research stage synthesizes sources into key facts, data points, and narrative angles
4. Outline stage structures the piece with sections and assigned evidence
5. Draft stage writes the full article using your style fingerprint, persona, and past performance insights
6. 4 review agents score independently — editing, fact-checking, engagement, and AI detection — each enforcing minimum quality thresholds
7. Polish loop revises and re-reviews up to 3 times; content that can't pass the quality gate is never published
8. Adapt stage fans out to platform-native formats (Markdown for blogs, threaded tweets, podcast scripts with TTS)
9. Analytics sync pulls engagement metrics back in, so the next run learns from what resonated

---

Quality: The Anti-Slop Stack

Most AI content fails because there's no quality gate between generation and publish. Ghostwriter's review layer is the core differentiator:

• 93-phrase blacklist — "delve", "it's important to note", "in today's rapidly evolving landscape", and 90 more are hard-rejected
• 4 parallel review agents — Editor (structure, readability, voice match), Fact Checker (accuracy, source coverage), Engagement (hook strength, reader pull-through), AI Detection (naturalness, sentence-length variance, burstiness)
• 9 scored dimensions with configurable minimums per channel
• Revise-or-reject loop — fails get polished and re-reviewed up to 3 times. If it still doesn't pass, it goes to a dead-letter queue instead of your audience

---

Style Fingerprinting

Point Ghostwriter at your best writing and it extracts a quantitative style profile — no LLM calls, no fine-tuning, zero cost:

ghostwriter fingerprint https://your-blog.com/favorite-post

What it captures: sentence length distribution, paragraph variation, contraction frequency, formality, vocabulary patterns, use of headings/lists, greeting and signoff style, top n-grams. This profile is injected into the draft prompt so generated content matches your voice.

---

Install

npm install -g @npow/ghostwriter

Quick Start

Prerequisites: Node.js >= 22, pnpm, Docker (for Postgres + Redis + Temporal)

# Clone and install
git clone https://github.com/npow/ghostwriter.git && cd ghostwriter
pnpm install

# Start infrastructure
docker compose up -d

# Configure
cp .env.example .env
# Edit .env with your API keys (at minimum: ANTHROPIC_API_KEY)

# Build
pnpm turbo build

# Connect a publishing platform
ghostwriter connect wordpress-com   # OAuth flow — opens browser
ghostwriter connect wordpress       # Self-hosted — Application Passwords
ghostwriter connect twitter

# Create a channel from a description
ghostwriter create "a daily recipe blog with Mediterranean focus"

# Or scaffold manually
ghostwriter init my-channel
ghostwriter validate my-channel
ghostwriter run my-channel

---

Example Channels

Ghostwriter ships with three example channels to show the range:

Channel	What it does	Data sources	Platforms	Schedule
Easy Weeknight Meals	30-min recipe blog in a casual parent's voice	Spoonacular API, Budget Bytes RSS	WordPress	Tue/Thu 9am
Tech Digest Podcast	5-min daily tech news recap, skeptical tone	Hacker News, TechCrunch, The Verge RSS	Podcast (Buzzsprout) + Twitter	Weekdays 7am
Weekly Stock Recap	S&P 500 analysis, no-jargon style	Polygon API, Yahoo Finance RSS	WordPress + Twitter	Sat 10am

Each channel config defines voice (persona, opinions, verbal tics, vocabulary), data sources, quality thresholds, and publish targets — all in one YAML file.

---

Architecture

Monorepo with 10 packages, orchestrated by Turborepo and Temporal.

packages/
  core/                 # Schemas, config, logger, anti-slop blacklist, connections store
  cli/                  # Commander-based CLI (create, connect, run, fingerprint, ...)
  data-ingestion/       # RSS feeds, APIs (Polygon, Spoonacular), caching & dedup
  style-fingerprint/    # Writing style extraction — pure computation, no LLM
  content-pipeline/     # Multi-stage LLM pipeline: research → outline → draft → review
  publishing/           # Platform adapters: WordPress, Twitter, Buzzsprout podcasts
  site-setup/           # WordPress.com site provisioning + OAuth
  database/             # Drizzle ORM + PostgreSQL (channels, runs, artifacts, publications)
  orchestrator/         # Temporal workflows for scheduled content generation
  monitoring/           # Analytics sync, performance insights, engagement feedback

---

Environment Variables

See .env.example for the full list. The only required key to get started is ANTHROPIC_API_KEY.

Development

pnpm turbo build       # Build all packages
pnpm turbo test        # Run tests
pnpm turbo dev         # Watch mode

Install from npm

npm install -g @npow/ghostwriter
ghostwriter create "a weekly tech blog about AI safety"

License

MIT