The Style Sheet: Keeping Names, Terms and Numbers Consistent with AI

Most inconsistencies in a book are failures of memory, not of knowledge. You spelled a name one way in chapter two and another way in chapter nine. The wider workflow is in the complete map of editing with AI; this piece digs into one problem: consistency across a long manuscript.

What a Style Sheet Is

A style sheet is a decision log, not a rulebook. A house style guide says what is generally correct in the language; the sheet says what you have chosen in this manuscript. Many consistency decisions have no right answer: both ebook and e-book are fine. The question is never which one is correct, only which one your book uses.

When the writer and the editor are the same person, memory is the weakest link. The sheet drags those decisions onto the page. Every line has three parts:

• The item: the thing being decided — a name, a term, a way of setting numbers.
• The chosen form: the only form that appears in the text from now on.
• A note: the rejected forms, the reason, and the first place the item shows up.

A sheet that grows into an essay never gets read; three columns is enough.

What Belongs in It

Do not start with a wishlist, but with whatever actually repeats in your text:

• Proper nouns: people, places, institutions, titles of works. The hard part is rarely the original spelling; it is the form you use in your own language.
• Terminology: where one concept has both a native coinage and a loanword — user interface or UI. The call is made in chapter one and holds to the appendix.
• Quotations, punctuation, headings: which quotation marks, where punctuation sits against the closing mark, how titles and subheads are set.
• Your own high-frequency words: what you call the object in your hands — book or manuscript; chapter or part.
• Numbers, dates and units: the next section.

Building a Style Sheet from Your Own Text with AI

You already have a partial style sheet; it lives in your habits rather than in a file. The model can pull it out.

Pattern one: extract the vocabulary

Feed the manuscript chapter by chapter and ask for a table of proper nouns and terms with a count for every variant. The critical instruction is the constraint: list only, do not change the text. Leave the model free and it starts rewriting, and the evidence you need is gone.

Pattern two: surface the competing forms

Ask which items appear in more than one form and how often. Find-and-replace never catches a variant that differs by one letter; that is where the model earns its place.

Pattern three: test the sheet against the text

Hand the sheet over with a single chapter and ask for violations only: which lines conflict with which entries. A report, not a repair.

Language models produce plausible errors: a form that never appeared in your text, a location where the string is not there. Every row is a candidate, not a fact: confirmed by search or thrown away. The model cannot know which variant you kept on purpose.

Finding Inconsistencies

Two directions, and you need both. Top-down: sheet in hand, you flag every occurrence of a non-chosen form. Bottom-up: the model surfaces competing forms and you decide which one goes into the sheet.

Before the sweep starts, freeze the sheet. Keep taking decisions mid-sweep and you will spend the project chasing your own rulings. What you want is a table, not prose:

Strike a wrong location, keep the decision: the location column is for tracking only. Some inconsistencies are surface-level rather than lexical — spacing, letter forms, spaces around punctuation — and those do not mix into vocabulary decisions; that family gets its own treatment in copyediting Persian text with AI.

Numbers, Dates and Units

Numbers attract the most decisions and the least attention. Settle these once:

• Digit shapes: in a Persian-set book the body uses Persian digits; Latin digits only in URLs, codes and standard numbers.
• Separators: Persian uses distinct characters for thousands and decimals, not the Latin comma and full stop. Record them explicitly, even if the choice is none.
• Calendar, ranges, units: one calendar or two, with a fixed order and bracket style; one style for ranges and abbreviations; one rule for the space before a unit.
• Cross-references: how chapters, figures and tables are named in running text. Those survive proofreading easily, because each one looks right in its own paragraph.

Keeping the Sheet Alive Through the Project

A style sheet is not a snapshot but a living document. Add decisions the same day you make them; one postponed to tomorrow is lost.

• One file next to the manuscript, not ten scattered notes.
• Date every entry and mark it settled or open; do not apply an open entry yet.
• If someone else edits or proofreads later, the sheet is the handover document.

The sheet normalises the mechanical layer only — names, terms, numbers, marks. Rhythm and word choice are not its business. If normalisation starts sanding down your voice, the line has been crossed; holding on to it is covered in keeping your voice when editing with AI.

1. Extract the repeated vocabulary from the text.
2. Record the chosen form for each item.
3. Freeze the sheet.
4. Sweep the manuscript and log the violations in a table.
5. Fix the text, not the sheet — unless you recorded a new decision.
6. Hand the sheet over with the manuscript.