Imagine a world where every author has a writing partner who never gets tired, never breaks the voice, and never leaves until the chapter is done.

GhostAI is an MIT-licensed toolkit of AI skills that runs inside Claude Code and Codex. You install it on your own laptop. It works on your own manuscript. The AI calls go through your own agent account.

Most AI writing tools generate slop. GhostAI keeps your voice. It reads your full manuscript structure, maintains context across chapters, and edits in place — no copy-paste, no browser window. The AI works where you work.

GhostAI is built on three design principles.

The AI suggests. You decide. Every skill has an approval gate. /ghost-start shows you its topic sharpening and asks before it scaffolds. /ghost-draft proposes a section-by-section plan and waits for your sign-off before writing any prose. /ghost-rewrite shows you structural skeletons — not prose — so you pick the angle before a single sentence is written. /ghost-edit presents every change as a suggestion you accept or reject, like a real editor in track changes. Nothing is auto-applied. Nothing ships without your yes.

Here's what that looks like in practice. You run /ghost-start and type your topic:

"GhostAI."

It pushes back:

"That's a tool, not a book. What's the ONE thing your reader should walk away able to do?"

You try again. It offers sharper versions. You pick one. The same pattern repeats for audience — "developers" gets rejected because it's not a person. By the time the interview is done, you've landed on a topic, an audience, and an angle that are specific enough to write to. The sharpening is the value, not the generation.

Diagnose before you act. GhostAI never jumps straight to output. /ghost-rewrite maps the current structure of a passage paragraph by paragraph, then articulates what's actually wrong before proposing alternatives. /ghost-review reads the full manuscript before surfacing a single finding. /ghost-outline challenges every chapter to earn its place before suggesting changes. The diagnosis is often more valuable than the fix — once you can see the shape of the problem, you frequently know the answer yourself.

Your voice is non-negotiable. Every skill reads your voice profile on every run. /ghost-edit runs a voice-consistency check on every pass. /ghost-interview preserves your verbatim phrasing. /ghost-rewrite tags every sentence as verbatim, rephrased, or new so you can see exactly what changed. Slop detection is advisory — flagged, never auto-fixed. The whole system is designed so that the more you use it, the more it sounds like you, not less.

Pick the tab that matches your agent. The installer creates symlinks so your agent discovers each /ghost-* skill. No build step, no binaries, no Python or Node dependencies.

git clone git@github.com:leanpub/ghostai.git ~/.claude/skills/ghostai
~/.claude/skills/ghostai/install.sh --claude

Verify: open Claude Code in any directory and run /ghost-status.

git clone git@github.com:leanpub/ghostai.git ~/.codex/skills/ghostai
~/.codex/skills/ghostai/install.sh --codex

Verify: restart Codex, then run "use ghost-status".

git clone git@github.com:leanpub/ghostai.git ~/.claude/skills/ghostai
~/.claude/skills/ghostai/install.sh --all

Verify: run /ghost-status in Claude Code, then restart Codex.

HTTPS clone: replace git@github.com: with https://github.com/ if you don't have SSH keys set up.

GhostAI checks for updates once per day when you run any /ghost-* command. When an update is available, it'll let you know. To update:

# Option 1: Use the skill
/ghost-update

# Option 2: Run the script directly
cd <path-to-ghostai-repo> && git pull && ./install.sh --all

• Claude Code and/or Codex
• An API key or account for your chosen agent
• macOS, Linux, or WSL2 on Windows
• A manuscript in Markua format (or run /ghost-start to create one)

Three gotchas account for most install problems:

1. Corporate proxies block the git clone. Ask your network admin to whitelist github.com, or clone from a personal network.
2. Missing SSH keys. Use the HTTPS clone URL — no keys needed.
3. Windows on plain PowerShell. Install WSL2 and run the Linux commands from inside it. PowerShell can host Claude Code, but GhostAI assumes a POSIX shell.

The whole toolkit hangs off The Loop — a repeatable cycle that takes a chapter from idea to publish. Get this and you've got the spine of GhostAI.

1. Start once. Define the book — topic, audience, voice. This happens once per book, not once per chapter. → /ghost-start
2. Outline each chapter. Push your structure past your first instinct. → /ghost-outline
3. Pick your path. Three doors out of the empty file — choose the one that matches your state. → /ghost-draft, /ghost-expand, or /ghost-interview
4. Edit, or rewrite. Polish line by line with /ghost-edit — that's where most of your time will go. Escalate to /ghost-rewrite when a section is structurally wrong, not just rough.
5. Review when ready. Once you have three or more chapters, take the cross-cutting view. → /ghost-review
6. Publish. Generate a Leanpub preview, ship it, repeat.

Steps 2–4 are the chapter loop — you go around them for every chapter. Steps 5–6 are the manuscript loop. Step 1 is the only one you do once.

                   /ghost-start
                         │
                         ├──optional──→  /ghost-voice
                         │                     │
                         │                     ▼
                         │               /ghost-train
                         │                     │
                         ├◄────────────────────┘
                         ▼
                  /ghost-outline
                         │
                         ▼
          ┌──────────────┼──────────────┐
          │              │              │
     /ghost-draft   /ghost-expand   /ghost-interview
     "blank page,   "I have notes  "I know this in my head
     tell me what    but they      but I can't get it on
     it could be"    won't grow"   the page in my voice"
          │              │              │
          └──────────────┴──────────────┘
                         │
                         ▼
              /ghost-edit  ⇄  /ghost-rewrite
              (line-level)     (section-level —
                                when polish isn't enough)
                         │
                         ▼
                /ghost-review  (after 3+ chapters)
                         │
                         ▼
                      publish

/ghost-status — the heartbeat. Run between every step.

The hardest moment in The Loop is the empty chapter file. GhostAI gives you three doors out of it. They're not sequential steps — they're alternates. Pick the one that matches the state you're actually in.

/ghost-draft — when you have a blank page. A chapter title, maybe some outline notes, and nothing else. /ghost-draft interviews you about what the chapter should do, then generates a first draft you can react to. The AI proposes. You react. The draft doesn't have to be good. It has to exist.

/ghost-expand — when you have notes that won't grow. Bullets in a notes app, outline points, a half-paragraph you wrote on the train. The text is there but it won't turn into prose. /ghost-expand reads what you have, asks for the missing connective tissue, and turns the bullets into paragraphs. You wrote the bones. The AI grows the meat.

/ghost-interview — when you know it cold but the prose won't sound right. You can talk about it for an hour at a meetup, answer any question a colleague asks. But every time you try to write it down, it sounds like someone else. /ghost-interview flips the script: the AI asks short questions, you answer in your own words (typed, or dictated through speech-to-text like Wispr Flow), and the AI weaves what you actually said into prose. The words are literally yours.

You can use a different door for every chapter. The chapter you draft today, you might expand tomorrow, and interview the day after that. If a skill produces something you don't recognize as your voice, stop — don't fight the prose. Run /ghost-interview instead. The interview path is the strongest voice match because the words are literally yours.

Writing is rewriting. The first draft of every chapter is a punching bag. The second pass starts to feel like writing. The third pass is where the chapter becomes good. By the tenth pass, the chapter is finally yours.

Here's what a writing morning looks like with GhostAI:

1. Open Claude Code and run /ghost-status. See exactly what you finished yesterday and which chapter is empty next.
2. Pick your door. Blank file → /ghost-draft. Bullets that won't grow → /ghost-expand. Know it cold but it won't sound right → /ghost-interview.
3. Run /ghost-edit. Polish until you'd ship it. If a section keeps not working line by line, escalate once with /ghost-rewrite.
4. Run /ghost-status again. Word count up, voice profile holding, one more chapter polished.

That's it. The Loop is the toolkit, repeated. Every morning. Every chapter.

/ghost-start — The topic interview. Run it once, at the very beginning, when you have a book idea but not yet a book. It challenges your topic ("Developers" isn't an audience. "AI" isn't a topic.), generates a starter outline, scaffolds the manuscript, and extracts a voice profile from a writing sample you provide. That voice profile is what keeps everything sounding like you instead of like a model.

/ghost-outline — Deep structure work. Goes beyond the starter outline from /ghost-start: challenges every chapter to earn its place, finds the chapter your reader needs that you didn't think of, and catches symmetry problems (real books have heavy chapters and light chapters). Run it every time you sit down to think about your book's structure.

/ghost-draft — Blank page to first draft. Interviews you about what the chapter should do, proposes a section-by-section plan, gets your approval on each approach before writing any prose, then drafts it.

/ghost-expand — Bullets to prose. Takes rough notes, outline points, or stream-of-consciousness jotting and grows them into finished paragraphs. Preserves your intent and phrasing where it works, smooths what doesn't.

/ghost-interview — Talk your chapter into existence. The AI asks short, voice-input-friendly questions. You answer in your own words. The transcript becomes prose with your verbatim phrasing leading. The strongest voice match of the three doors, because the words are literally yours.

/ghost-edit — The daily driver. The command you'll run more than any other. A copy editor for one chapter at a time that runs six specialist passes: grammar, clarity, tone and voice consistency, technical accuracy, advisory AI slop detection (flagged, never auto-fixed), and a pre-flight publish checklist. Every change is a suggestion you accept or reject — like a real editor working in track changes.

/ghost-rewrite — Section-level structural revision. For when a passage exists but the staging is wrong — unclear POV, buried lead, wrong order of ideas. Diagnoses the structural bug first (the value is in articulating what's actually wrong), then proposes two structurally different rewrites so you pick the angle rather than nudge your way into it. If you find yourself reaching for /ghost-rewrite more than once a chapter, your outline is probably the real problem — go back to /ghost-outline.

/ghost-review — The development editor. Reads your entire manuscript, not just a chapter. Catches the problems no individual /ghost-edit pass could catch: a concept introduced in chapter 2 but defined in chapter 5, an argument that stalls in the middle, the same point repeated three times across different chapters, pacing that drags or rushes. This is where GhostAI's cross-chapter memory matters most. Three triggers for running it: you just finished your third chapter, you're about to publish a preview, or you haven't written in two weeks and you're picking the book back up.

/ghost-status — The heartbeat. A dashboard for your book at a glance: word counts per chapter, estimated reading time, chapter balance, drafted versus outline. Fast, free, and it tells you exactly where the book is. Run it between every step of The Loop — the first command tomorrow and the last command tonight.

/ghost-voice — Manage your voice profile across machines, projects, and co-authors. Update it from a fresh writing sample, promote it from your laptop into the book's git repo so co-authors share the same voice, or copy it across machines.

/ghost-train — Teach GhostAI what you think about your own writing. Three short interviews: what you're good at, where you struggle, and the patterns you can't stand seeing on your page. Answers are synthesized into your style guide and learnings so every other skill applies them automatically. Run it during onboarding, after a tough edit session, or any time you want Ghost to internalize how you actually think about your craft.

/ghost-update — Updates GhostAI to the latest version. Pulls the latest code and re-creates skill symlinks. GhostAI also checks for updates automatically once per day when you run any /ghost-* command, and will let you know when a new version is available. origin/main

Most AI writing tools are stateless — every session starts from scratch. GhostAI maintains six persistent artifacts that accumulate across sessions, so the toolkit gets smarter about your book and your voice the more you use it.

Your extracted writing style: sentence length distribution, vocabulary level, tone, person, formality, signature phrases, em-dash density. Created by /ghost-start from a writing sample you provide (a blog post, an email, a chapter you've already written). Every skill reads this file on every run — it's the primary mechanism that keeps generated and edited prose sounding like you instead of like a model.

Update it from a fresh sample with /ghost-voice. The profile improves as you feed it more of your writing.

A human-readable document of your stated preferences — what you're good at, where you struggle, and the patterns you actively dislike on your page. Created by /ghost-train through three short interviews. Skills use it as a soft constraint: lean into stated strengths, flag stated weaknesses harder, never generate stated dislikes.

Unlike the voice profile (which is extracted from samples), the style guide captures what you think about your own writing. Both matter. They complement each other.

A running log of terminology decisions, style preferences, and craft insights — one JSON entry per line, accumulated across sessions. When /ghost-edit discovers you prefer "dataset" over "data set", it appends an entry. When /ghost-rewrite learns you prefer cold opens over setup paragraphs, it logs that. When /ghost-train records that you dislike rhetorical questions, that goes in too.

Every skill loads the relevant learnings on startup. Types include:

The learnings file is append-only and grows over the life of your book. It's the primary mechanism for GhostAI getting more opinionated in your direction over time.

Timestamped reports from /ghost-review — each one a full-manuscript development edit capturing structural issues, repetition, tonal drift, gaps, and cross-chapter contradictions. Saved so you can track how the manuscript evolved and so future reviews can reference prior findings.

Detailed records from /ghost-rewrite — the original passage, the structural diagnosis, the candidate skeletons considered, the chosen skeleton, the filled-in prose with provenance markers ([v] verbatim, [r] rephrased, [n] new), and the consistency sweep results. Future runs on the same passage read prior logs to avoid repeating strategies that already failed. /ghost-review can surface "passages with multiple rewrites" as a signal that something deeper is wrong.

Saved by /ghost-interview alongside the manuscript. Each transcript records the questions asked, your verbatim answers, and any follow-up exchanges. Transcripts are reusable: you can re-synthesize prose from them later, feed them into your voice profile as writing samples, or reference them when the same topic comes up in another chapter.

Every artifact lives in one of three tiers. The highest existing tier wins on load (no merging):

GhostAI asks where to save when you first create a profile (during /ghost-start or /ghost-voice), and remembers your choice via a .tier file so subsequent silent writes (learnings, reviews, rewrite logs) land in the same tier without re-prompting.

Picking a tier:

• One book, one machine → project-local (the default)
• One book, multiple machines → in-repo (syncs through git)
• Many books, consistent voice → global
• Co-authored book → in-repo for voice and style guide; personal learnings can stay project-local

Run /ghost-voice to update, promote between tiers, or share with co-authors. See shared/config-hierarchy.md for the full model.

GhostAI expects Markua format (Markdown for books), the native format for Leanpub. Books follow the default Leanpub template:

my-book/
  manuscript/
    Book.txt          # Chapter manifest (one filename per line)
    introduction.md   # Chapters in Markua
    chapter-01.md
    chapter-02.md
    resources/        # Images and code samples

Book.txt lists chapters in reading order:

{frontmatter}
introduction.md
{mainmatter}
chapter-01.md
chapter-02.md
{backmatter}
appendix.md

Just use this with a Claude Pro or Claude Max plan. Seriously.

Every GhostAI skill runs through Claude Code. On a Claude Pro or Claude Max plan, your usage is already covered — the skills cost you nothing extra beyond the flat monthly subscription. No per-token billing, no surprise invoices, no watching a meter tick while you edit your book. A single /ghost-review pass on a full book can burn through what a pay-as-you-go API user would consider a meaningful bill. On a Claude plan it's just included.

(Note: these are Anthropic's Claude Pro / Claude Max subscription plans, not Leanpub's author plans — unrelated products that share tier names.)

If you're using the Anthropic API directly, rough costs for a 50,000-word manuscript on Claude Sonnet:

Run /ghost-status for estimates specific to your manuscript.

GhostAI uses the universal SKILL.md format. While built and tested with Claude Code, the same skill files work with other SKILL.md-compatible agents:

• Claude Code (primary, fully tested)
• Cursor
• Gemini CLI
• Codex CLI

Some interactive features may be simplified on non-Claude agents. The core writing intelligence works everywhere.

ghostai/
  ghost-start/SKILL.md         # Book onboarding wizard
  ghost-outline/SKILL.md       # Structured outline builder
  ghost-review/SKILL.md        # Development editor
  ghost-edit/SKILL.md           # Copy editor
  ghost-draft/SKILL.md          # First draft generator
  ghost-expand/SKILL.md         # Bullet-to-prose expander
  ghost-interview/SKILL.md      # Interview-driven writing
  ghost-rewrite/SKILL.md        # Section-level structural revision
  ghost-status/SKILL.md         # Manuscript dashboard
  ghost-train/SKILL.md          # Author self-assessment
  ghost-voice/SKILL.md          # Voice profile manager
  ghost-update/SKILL.md         # Self-updater origin/main
  shared/                       # Shared skill infrastructure
    preamble-core.md            # Manuscript detection & loading
    resolve-config.sh           # 3-tier config resolver
    config-hierarchy.md         # Author-facing tier model
    markua.md                   # Markua syntax cheat sheet
    anti-slop.md                # AI writing pattern detection
    voice.md                    # Editorial voice guidelines
    leanpub.md                  # Leanpub conventions
    output-format.md            # Terminal output formatting
    interaction-states.md       # Error/empty state handling
    update-check.sh             # Daily update checker (sourced by preamble)
  install.sh                    # Installer (Claude Code, Codex, or both)
  update.sh                     # Updater (pull + re-link)
  CLAUDE.md                     # Skill routing rules
  README.md                     # This file

MIT

Found a bug? Have a feature request? File an issue or talk to us during your Book Weekend or Boot Camp session.