Ghostwriter

A ghostwriter is a professional writer hired to draft or edit literary works, articles, or speeches that are published under another person's name.

An agent skill set that writes text in your own voice instead of generic AI prose, with a check and polish pass built into the pipeline. Works with Claude Code, Codex CLI, ChatGPT desktop, OpenCode, and any other agent that reads skills off the filesystem.

Installing

npx skills add an2n/ghostwriter --full-depth --all

--full-depth is required since this repo has three skills nested under skills/*/ rather than a single SKILL.md at the root. Drop --all to pick specific skills or agents interactively, or add -g for a global (user-level) install instead of project-level. See npx skills --help for the full option set.

Restart your agent (or start a new session) and the three skills are available - try /ghost (Claude Code), or just say "rewrite this in my voice" with some text.

For agents without filesystem-based skill folders (ChatGPT web, Gemini, Cursor, Aider), there's nothing to install - paste the contents of skills/ghost/SKILL.md (and the other two, if you want the full pipeline) directly into the conversation.

As a Claude Code plugin:

/plugin marketplace add an2n/ghostwriter
/plugin install ghostwriter@ghostwriter

Installed this way, Claude Code namespaces skills by plugin name, so ghost, check, and polish show up as ghostwriter:ghost, ghostwriter:check, and ghostwriter:polish - same skills, just addressed through the plugin name.

As a Codex CLI plugin:

/plugins

Add an2n/ghostwriter as a marketplace from the plugin browser, then install ghostwriter from it. Start a new session afterward for the bundled skills to load.

What's in here

Three skills, each with one job:

• ghost — the orchestrator. Loads your voice profile, runs the pipeline below, and drafts the actual rewrite.
• check — find-only. Scores a piece of text across four tell clusters (word choice, rhythm, document shape, voice and construction) and returns a verdict from "reads human" to "heavily AI" based on weighted tell density. Never rewrites anything.
• polish — rewrite-only. Applies six moves (word choice, sentence rhythm, cutting padding and imposed structure, anchoring specifics, letting a voice through, punctuation) to move text in the human direction. Never scores anything.

ghost is the one you actually invoke. check and polish are also directly callable on their own if you just want a score or just want a cleanup pass without the voice-profile layer.

How the pipeline works

Say "rewrite this", "write this better", "reply to this", "in my voice", or invoke /ghost with text:

1. Load the voice profile from ~/.claude/ghostwriter-profile.md (see below). If it doesn't exist yet, it asks once for 2-3 writing samples and builds one.
2. Check the input — if there's existing text to rewrite, check runs on it first to find out what's actually wrong, so the rewrite targets real problems instead of applying every move blindly.
3. Clean — polish runs the hygiene pass, directed by what the check found.
4. Apply the voice profile on top, overriding polish's generic defaults wherever they conflict with how you actually write.
5. Verify the output — check runs again on the draft. If it doesn't read as human or mostly human, the pipeline goes back to polish and redrafts, capped at two cycles.
6. Return the text, with a one-line plain-text status after each step as it happens (no leading symbols, no emoji, no batching everything at the end) — never the full check report inline.

A clean check result means no known rule-based pattern fired. It is not proof the text would pass a real trained classifier (Pangram, GPTZero) — check is a self-graded rubric, not a substitute for one.

The voice profile

ghost needs ~/.claude/ghostwriter-profile.md to know what "your voice" actually means. This file is personal, machine-local data — it never lives in this repo or gets committed anywhere. It's built once from 2-3 real writing samples (a LinkedIn post, an email, a Slack message — whatever's natural) and covers sentence shape, real vocabulary, punctuation habits, and register range across the contexts you actually write in.

If the profile drifts from how you actually write (you keep correcting the same thing), ghost will offer to update it — never silently, and never off a single correction.

Attribution

polish and check build on the functional idea of pairing a multi-category AI-detection scorer with a matching multi-lever rewrite pass, an approach explored by the humanize project by Harshaneel Gokhale (MIT License) and, for the underlying pattern knowledge, by Wikipedia's "Signs of AI writing" (CC BY-SA 4.0, maintained by WikiProject AI Cleanup) and the humanizer skill (MIT) that also draws on it. The clustering scheme, scoring mechanic, rewrite protocol, and wording in both skills here are their own, rewritten to be independent of any one source's specific structure or phrasing rather than a copy of it.

This repo is MIT licensed (see LICENSE).