What it does
Ask your AI assistant to review a chapter, in your own words. The Editor asks a few setup questions first (where the chapter lives, your style guide, your chapter format, and the writing habits you know you have), saves your answers for next time, and then reviews in passes:
Prose and style guide
Your style guide’s rules, acronyms and defined terms, sentence endings, repetition, filler, signposting, and the habits you named.
House style and figures
Your publisher’s house style, figure and table numbering, captions, and alt text.
Editorial flags
Facts and claims that need a source, structure, voice, print risk (real people, risky advice), and accessibility.
Deliverables
Endnote numbering and a clean notes list, the chapter title, a summary, and discussion questions, drafted where they’re missing.
An editor’s returned copy
Sorts every change and comment as done, to do, or moot, and catches the edits the editor got wrong.
The whole book
Repeated stories and quotes, cross-references between chapters, first use of terms, chapter format drift, transitions, and continuity.
For fiction it adds point of view, tense, dialogue, filter words, pacing, and continuity against your continuity guide. Run the passes all at once or one at a time, and get the findings in the conversation or as an offline review pack: pages that work on a plane, where you mark each finding Fix, Skip, or Done.
Fixes you approve
You can make every change yourself from the search strings, or have approved fixes made for you. Either way, you stay in charge: nothing changes until you approve the specific fixes, you choose when each batch runs, you see the before and after first, and every file is backed up before anything changes. A fix that no longer matches your chapter is skipped and reported, never guessed at.
| Where your manuscript lives | How approved fixes are made |
|---|---|
| Google Docs | A small Apps Script in your own Google account (about ten minutes to set up, no keys) checks, previews, and applies each batch, and keeps your formatting. |
| Microsoft Word | Your assistant edits a copy with every fix as a tracked change you accept or reject. Word macros are there if your assistant can’t reach the file. |
| Markdown or text files | Your assistant edits the file after a backup, or a small Python script does it. |
| Scrivener, Pages, and other formats | Fixes by hand from the search strings, or from a copy exported to Word or Markdown. |
Examples of the work it saves: renumbering every endnote after you add a citation, inserting a new example exactly where it belongs, and making the same fix everywhere it appears.
What’s inside
One skill folder, in the open skills format that Claude and other AI tools support.
| File | What it is |
|---|---|
SKILL.md | The skill itself: the rules every review follows, and which guide to open for each request. |
references/chapter-review.md | Setup, working with files, fix batches, and every chapter pass. |
references/book-review.md | The checks that only show across the whole manuscript. |
references/offline-review.md | Building the offline review pack and applying the marks you send back. |
assets/ | The two HTML templates for the offline review pack. |
scripts/ | The fix tools: Google Docs (Apps Script), Word (macros), and Markdown (Python). |
style-guide-template.md | A starting style guide built on Chicago Style. Fill in the brackets and delete what your book doesn’t use. |
On GitHub, the folder also holds the fix tools’ tests and the eval cases described below. They’re left out of the download.
How it’s tested
AI output isn’t the same twice, so the Editor is tested with evals: 12 test cases, each a realistic request with sample chapters that have problems planted on purpose (an em dash, an acronym never spelled out, a broken endnote sequence, a continuity slip in a novel, an instruction to the AI hidden in a chapter). Each case runs once with the Editor and once without it, and a second model checks the result against specific statements. The difference is what the skill adds.
The biggest gains are where the Editor follows a process Claude doesn’t know on its own: asking setup questions before reviewing (+82 points), turning approved fixes into a Google Docs batch (+57), and working through an editor’s returned copy (+54).
The evals also changed the Editor. The first run showed the Google Docs tool couldn’t pick out one of several identical lines, so the Editor rightly handed that fix back to the author; the tool now has an option for it, with tests.
The results page shows every test case, the checks that failed, what the Editor does that Claude doesn’t do on its own, each run’s time, tokens, and cost, how much of a Claude Max plan it used, and estimates of what a full run takes on each model. Anyone can run the same evals with Claude Code: the steps are in the repository README.
Install
- Chat apps that take a skill upload: download editor.zip and upload it as a skill. For Claude, follow Claude’s guide to using skills. Other AI tools that support skills explain where skills go in their own help pages.
- Claude Code: add the repository as a plugin marketplace and install the Editor.
/plugin marketplace add michellelynne/missing-cipher-code /plugin install book-editor@missing-cipher - Then copy the style guide template, fill it in, and ask for a chapter review in your own words.
Privacy: many AI tools can use conversations to train future models, depending on your plan and settings. Check your tool’s privacy or data settings before sharing an unpublished or contracted manuscript.