Which editing pass should come first?
Reading from page one and fixing whatever catches your eye mixes incompatible jobs. You may rewrite a paragraph whose chapter is later removed, correct punctuation around an unsupported claim, or miss that six chapters end with the same conclusion expressed in different words.
A pass gives the editor one question and one stopping rule. Structure asks whether the book is the right book. Evidence asks whether readers can trust important assertions. The line edit asks whether the accepted meaning reaches the reader clearly. Keeping those questions separate makes decisions easier to audit and hand off.
Freeze a review copy
Preserve the manuscript you received, duplicate it for editing, and record the date, version, known sources, intended audience, destination, and decision-maker. Without a stable baseline, later reviewers cannot distinguish an original defect from an editorial change.
Example: an editing log for three repeating chapters
Suppose chapters two, four, and six all tell freelancers to “set clear expectations.” The words differ, but each section repeats the same advice. Search alone may miss it. Summarize what the reader learns in each section and compare the summaries.
| Location | Problem | Edit | Acceptance check |
|---|---|---|---|
| Chapter 2 | The main explanation lacks an example. | Keep the definition here and add a completed project-scope sheet. | The reader can name deliverables, exclusions, and an approver. |
| Chapter 4 | The introduction repeats chapter 2. | Replace the recap with a client request that changes the agreed scope. | The reader uses the earlier sheet to decide what changes. |
| Chapter 6 | The conclusion repeats the same recommendation. | Link back to the sheet and ask the reader to check the final handoff against it. | The ending checks the promise instead of teaching it again. |
After making these changes, read the three sections together. Confirm that chapter four does not need a definition you accidentally removed and that chapter six points to the right artifact. The goal is progression; deleting every repeated word would not accomplish it.
Common AI writing problems and how to find them
| Symptom | Likely defect | Fast test | Do not start with |
|---|---|---|---|
| Chapters sound polished but interchangeable. | No distinct chapter jobs or progression. | Summarize each chapter's reader change in one sentence. | Rewriting introductions one by one. |
| The same advice appears under new headings. | Semantic repetition hidden by paraphrase. | Build a concept-by-chapter grid from a sample of three chapters. | Searching only for repeated phrases. |
| Claims feel specific, but references are thin. | Unsupported numbers, citations, or borrowed authority. | Trace five high-consequence claims to original sources. | Standardizing citation punctuation. |
| Advice is agreeable but hard to apply. | Missing actors, constraints, thresholds, and demonstrations. | Ask what a reader should physically do next and how success is observed. | Adding another list of general tips. |
| The book promises confidence everywhere. | Uncertainty and scope limits have been flattened. | Mark every universal, causal, and 'best' statement in one chapter. | Making the tone more authoritative. |
Estimate editing time from completed work, with separate unknowns
Choose three different passages: one that makes factual claims, one that gives instructions and one that shows the author's voice. Edit them to the intended publication standard, including checking their sources. Record time by activity, not just total words.
For example, suppose a hypothetical 1,000-word sample takes 25 minutes to restructure, 40 minutes to trace sources and 20 minutes to line edit. That is 85 minutes. At the same observed rate, 12,000 similar words would require 17 hours. This is an illustration of the calculation, not an industry estimate or a quote for your book.
Keep unresolved work outside that projection. A missing interview may take a day to obtain but only minutes to edit. A chapter that needs a new argument should be estimated as a rewrite. Sample a later chapter too: if its sources are weaker or it repeats earlier material, the first rate will not describe the whole manuscript.
- Ready for line editing: the structure works, the main claims are supported and examples answer the reader's questions.
- Needs substantive revision: the argument is useful but important steps, explanations or evidence are missing.
- Needs reconstruction: chapters cannot be given distinct jobs, or the central claim has no adequate support. Return to the outline or research rather than editing every sentence.
Seven editing passes for an AI-generated book
- Pass 01
Reconstruct the book
Method: Write a one-sentence job for every chapter without using its title. Place those jobs beside the reader promise and look for detours, missing prerequisites, duplicate jobs, and a conclusion the chapters have not earned.
- Pass 02
Compress repetition
Method: Track concepts, examples, definitions, cautions, and conclusions across chapters. Choose one primary home for each important idea; elsewhere, remove it, shorten it, or explicitly build on it.
- Pass 03
Audit claims and attribution
Method: Mark numbers, quotations, comparisons, causal claims, superlatives, medical/legal/financial implications, and statements about current platforms. Trace each to the original or rewrite it to match what is actually known.
- Pass 04
Replace generic examples
Method: For each major recommendation, add an observed case, a reproducible demonstration, or a clearly hypothetical scenario with named conditions. Include the difficult choice, not only the successful outcome.
- Pass 05
Restore the author's judgment
Method: Highlight recommendations, caveats, transitions, and conclusions. Ask who is making each judgment and on what basis. Rewrite interchangeable summaries with the author's reasoning, experience, limits, and criteria.
- Pass 06
Repair continuity and sentences
Method: Check terms, names, chronology, examples, promises, cross-references, and chapter handoffs before the line edit. Then cut throat-clearing, stacked qualifiers, empty conclusions, vague pronouns, and sentences that hide the actor or action.
- Pass 07
Proof the reader's artifact
Method: Inspect the exported edition outside the editor. Navigate from the table of contents, follow links, review image alternatives and captions, change screen or type size where possible, and inspect page behavior appropriate to the format.
How to remove repeated ideas across chapters
Phrase matching catches copied wording. It does not catch the same point regenerated with different nouns and sentence shapes. Build rows for the book's recurring concepts and columns for chapters, then note the job each appearance performs.
| Concept | Chapter 2 | Chapter 5 | Chapter 8 | Editorial decision |
|---|---|---|---|---|
| Start with one outcome | Full explanation and first example. | Explained again with a new metaphor. | Restated as a conclusion. | Keep the teaching in Chapter 2; turn Chapter 5 into a brief cross-reference; test the idea in Chapter 8 instead of restating it. |
| Interview the reader | Mentioned without a method. | Five-step method and sample questions. | No appearance. | Foreshadow the need in Chapter 2 and keep the usable method in Chapter 5. |
| Measure before changing | Presented as a universal rule. | Presented as a general best practice. | Case shows an exception. | Define the conditional rule once, then use the exception to refine it. |
Some repetition is necessary. A definition may need a short reminder, and a complex process may need a checkpoint. The map forces the editor to name the second appearance's job. 'The model said it differently' is not an editorial purpose.
Use a claim ledger that ends in a decision
A source list records what you read. A claim ledger records whether the manuscript is allowed to say a particular thing. Work at sentence level when a number, quotation, cause or current rule matters.
| Draft claim | Check | Decision |
|---|---|---|
| “A folder system increases productivity by 40%.” | No original research supplied; productivity is undefined. | Remove the number. Do not replace it with “research proves.” |
| “The coordinator introduced a map.” | The interview note says only that a map was proposed. | Change “introduced” to “proposed.” Check later records before describing a result. |
| “Readers can adjust text size in any EPUB.” | EPUB can also represent fixed-layout publications. | Limit the explanation to a reflowable edition and test the actual file. |
| A direct quotation with no recording or transcript location. | The wording cannot be matched to the original. | Hold it for verification or remove it. Do not silently turn a generated quote into testimony. |
For a real entry, add the chapter and paragraph, exact source location, check date and person who resolved it. Keep the old wording in the log so another reviewer can understand the correction. Link the published source where the reader needs it; an internal “verified” label is not a reader-facing citation.
When the author is the source, distinguish a personal observation from a general claim. “In my three workshops, participants asked for a file checklist” describes those workshops if the records support it. It does not establish what all workshop participants need.
A full paragraph edit: remove the claim, then rebuild the instruction
The following passage is an invented editing exercise, not a quoted customer manuscript.
Draft
“Research shows that organized digital files can increase productivity by 40%. In today's fast-paced world, a robust folder system is essential. Start by organizing everything into logical categories and adopt a consistent naming convention. This will eliminate confusion and allow your team to collaborate effortlessly. Remember to regularly review the system to ensure that it continues to meet your evolving needs.”
Editorial decisions
- Remove the 40% claim until its source is found. The passage supplies no study, population or definition of productivity. Adding a citation to an unrelated article would not repair it.
- Cut the opening commentary. “Fast-paced,” “robust” and “essential” supply no information about the actual file problem.
- Choose a specific reader task. Here, a project assistant needs to find the approved illustration without asking its designer.
- Replace absolute promises. A naming convention cannot guarantee that nobody uploads an old file. Include a way to identify the approved version and an owner for unresolved cases.
Revised passage
“Put the approved illustration in a folder named Approved, and keep drafts in a separate Working folder. Add a file index that names the asset, its intended use and the person who approved it. For example: garden-map-v3.png; visitor welcome sheet; approved by the event coordinator on 6 September. Before sending the folder, ask a colleague to find the approved map using only the index. If two files both appear current, resolve that conflict with the approver before the handoff.”
The filename, role and date are invented example details. In a real manuscript, use the author's actual process or label the workflow as a proposed method to test. The revised paragraph makes a modest promise: it explains how to identify one approved file and check that another person can find it.
What the line edit cannot decide
If the organization must use an existing document-management system, a new folder convention may be inappropriate. Ask that question before polishing this passage. The most fluent version can still be the wrong instruction for the reader.
How to check names, terms, and story continuity
Long drafts can change vocabulary, assumptions, chronology, and promises without producing an obvious local error. Record the canonical form of details that must remain stable, then check every occurrence.
- Terms: preferred name, first definition, allowed shorthand, and terms deliberately avoided.
- People and cases: spelling, role, pronouns, permissions, chronology, and what the example is allowed to prove.
- Numbers: unit, time period, calculation source, rounding rule, and every downstream use.
- Promises: outcomes introduced in the opening and where each is fulfilled or explicitly limited.
- Cross-references: chapter names, section destinations, figure numbers, links, and references to material that may have moved.
- Instructions: prerequisite, action order, expected result, exception, and recovery step when the result does not occur.
Why an AI detector score cannot replace editing
Grammar and style checks
Useful for possible mechanical defects and patterns. They cannot determine whether the sentence is true, necessary, fairly attributed, or appropriate for this reader.
Similarity searches
Useful for locating matching language for human review. A match still needs context, source, rights, quotation, and attribution analysis.
AI-origin detectors
Do not use a single score as proof of who authored a passage. Evaluate process records and the manuscript itself against the required standard.
Automated citation checks
Useful for broken identifiers and missing fields. The editor still has to open the source and confirm that it supports the exact claim in context.
Final manuscript and export checklist
| Area | Go when | Stop when |
|---|---|---|
| Purpose | The title, opening, chapters, and conclusion deliver the same bounded promise. | A surviving chapter cannot explain its contribution to that promise. |
| Evidence | High-risk claims have documented review and lower-risk facts were checked systematically. | A material quotation, citation, statistic, or current rule remains unresolved. |
| Authorship | The named author has supplied or approved the substantive examples, reasoning, limits, and recommendations. | No accountable person can explain or defend important advice. |
| Rights and rules | Permissions, licenses, disclosures, and destination requirements are documented. | Ownership, confidentiality, permission, or required disclosure is still unclear. |
| Artifact | The final file has working navigation, readable content, valid links, and format-appropriate behavior. | Only the editable source, not the reader-facing edition, has been reviewed. |
Sources for fact-checking and ebook validation
Editing can identify a decision, but the applicable platform or technical authority defines rules that may change. Record when those sources were checked and return to them immediately before release.
- NIST AI 600-1: Generative AI Profile: risk-management context for false or confidently stated content and other generative-AI risks.
- Amazon KDP Content Guidelines: current publisher responsibilities and platform content requirements.
- W3C EPUB 3.3: the normative specification when the reviewed delivery artifact is EPUB.
- W3C WAI image tutorials: purpose-based guidance for reviewing alternatives to informative, functional, complex, and decorative images.
AI book editing FAQs
How do you edit an AI-generated book?
Edit from the book level downward. First test the reader promise and chapter sequence; then review repetition, claims, examples, author judgment, continuity, sentences, and the exported artifact. Starting with grammar can polish material that should be cut or rewritten.
How much editing does an AI-assisted manuscript need?
There is no reliable percentage. The work depends on the sources, subject risk, instructions, model output, author contribution, and publication standard. Estimate from a representative sample and track defects by type instead of assuming a fixed amount of cleanup.
Can an AI detector prove who wrote a book?
A detector score should not be treated as proof of authorship or misconduct. Use production history, source notes, version records, interviews with the author, and direct manuscript review to evaluate how the work was made and whether it meets the required standard.
What should an editor fact-check first?
Prioritize claims that could cause harm or materially change a reader's decision: health, legal, financial, safety, technical, current-platform, numerical, causal, comparative, and attributed statements. Then review lower-risk factual details systematically.
Start with a chapter-by-chapter review map
Duplicate the manuscript, summarize each chapter's contribution, and mark any pair that performs the same job. That single map reveals whether the next action is restructuring, merging, or a closer evidence review. If the drafting process itself is still being designed, begin with how to write a book with AI. Before final production, use the EPUB, PDF, and DOCX guide and the ebook image review guide.