Why "revise this chapter" is a broken prompt
You paste in a chapter, type "make this better," and hit enter. What comes back reads smoother, sure. But three chapters later you notice the subplot thread you carefully planted is gone. The AI "improved" a paragraph right out of existence because it didn't know that sentence was load-bearing.
This is the core problem with one-shot revision prompts: you're asking a model to simultaneously judge structure, fix mechanics, and polish voice in a single pass, with no priority order. It has to guess what matters most. Sometimes it guesses fine. Often it optimizes for the thing that's easiest to fix — swapping "said" for "murmured," tightening a sentence — while leaving the actual problem (a scene with no stakes, a character who forgives too easily) completely untouched. Worse, a vague prompt gives the model permission to rewrite everything, which means it might fix your pacing issue by deleting the exact clause that made your protagonist's voice distinct.
The fix isn't a smarter prompt. It's a different structure entirely — one where each pass has exactly one job, and the jobs happen in an order where later passes can't undo earlier ones. I call this a revision ladder: you climb one rung at a time, and you don't move up until the rung below is solid.
This isn't a new idea in editing generally — professional editors have always separated developmental edits from copyedits from proofreading. What's different with AI is that the model has no instinct to keep these separate unless you force it. A human editor doing a line edit won't accidentally restructure your third act. An AI told to "polish this" might. So the ladder isn't just good practice — it's a guardrail against a specific AI failure mode.
If you want the more granular version of this idea, The Five-Pass Revision Order for AI-Assisted Novels breaks the process into five stages instead of three. This guide is the compressed, three-rung version — useful when you want the discipline of staged revision without managing five separate prompts per chapter.
Every revision pass should answer one question and be forbidden from touching anything else. The moment a prompt tries to fix structure and voice at once, you've lost control of what changes.
Rung 1: The structural pass — function, stakes, causality only
Before a single sentence gets polished, you need to know the scene is doing its job. Does it move the plot? Does the protagonist want something and fail to get it cleanly? Does what happens here cause what happens next, or could you delete the scene and nothing downstream would notice?
This is where most AI-drafted chapters actually fail — not at the sentence level, but at the scene-function level. A scene can be beautifully written and still be dead weight if it doesn't change anything. The structural pass exists to catch that before you waste time polishing prose that might get cut anyway.
The key move here is telling the AI, explicitly, to ignore prose quality. Don't let it comment on word choice or sentence rhythm at all. You want a diagnostic report, not a rewrite.
Read this chapter as a structural editor only. Do not comment on prose style, word choice, or sentence-level writing quality — I'll handle that in a later pass. For each scene in the chapter, answer these questions: (1) What does the POV character want in this scene, and what do they get instead? (2) What changes because of this scene — plot, relationship, or information — that wasn't true before it? (3) Does this scene's outcome cause the next scene to happen, or could the next scene occur unchanged if this one were deleted? (4) Where are the stakes unclear or too low to justify the scene's length? List findings scene-by-scene. Flag anything that reads as filler.
Why this works: you've boxed the model into a single lane. It can't dodge into "here's a nicer verb" territory because you've told it that's out of scope. You get a structural diagnosis you can actually act on — cut this scene, raise the stakes here, add a causal link there — before any prose gets touched.
If a scene fails this test, don't ask the AI to fix it inline. Go back to your outline or story bible and rework the scene's purpose first. If you drafted from an AI book outline, this is often where you discover the outline promised more conflict than the draft delivered — worth checking against the original beats before you touch prose. Tools like the Three-Act Beat Sheet are useful here for confirming a scene actually belongs where you put it.
One more thing worth checking at this rung: causality chains across chapters, not just within one. A common AI-draft failure is a character learning something in chapter 12 that should have consequences in chapter 14, but the model "forgot" by the time it got there. Ask specifically:
Here are chapters 11 through 15 of my thriller manuscript. Track one thing only: information the protagonist learns and whether later chapters act as though she knows it. List every instance where a character's later behavior is inconsistent with information they were given earlier, and note the chapter and paragraph where the inconsistency shows up. Do not suggest prose fixes — just flag the continuity breaks.
This kind of cross-chapter check pairs well with a dedicated character consistency pass if you're managing a large cast, since voice drift and knowledge drift tend to show up together.
Rung 2: The mechanics pass — sentence variety, tags, pacing
Only once the structure is locked do you move here. This rung is about the craft mechanics that make prose readable without touching voice: sentence length variation, dialogue tag repetition, paragraph pacing, and overused constructions. It's the pass that catches "she said" appearing four times in six lines, or six consecutive sentences all starting with the subject's name.
This is mechanical work, and it's exactly the kind of thing AI is good at spotting because it's pattern-based rather than judgment-based. The trap is letting the mechanics pass creep into rewriting for voice — which is rung 3's job. Scope it tightly.
This chapter's structure is finalized — do not suggest cutting, adding, or reordering scenes. Focus only on prose mechanics: (1) Flag any run of 3+ consecutive sentences with the same structure or opening word. (2) Flag dialogue tags that repeat the same verb within 10 lines of each other, and suggest which ones could become an action beat instead. (3) Identify paragraphs where sentence length doesn't vary — all long, or all short — and note where a rhythm break would help pacing. (4) Flag any paragraph longer than 200 words that could be split for pacing in an action scene. Give me a list with line references, not a rewritten chapter.
Notice the last instruction: ask for a list, not a rewrite. This is the single biggest thing that keeps rungs from colliding. If the AI hands you a fully rewritten chapter at the mechanics stage, you have no way of knowing whether it also changed voice, tone, or meaning along the way. A list of flagged issues lets you go in and fix each one by hand, or ask for a targeted fix on just that line.
Pacing deserves special mention here because it's the mechanics issue most likely to hide inside dialogue-heavy scenes. AI drafts — especially in romance novel with AI workflows and dialogue-driven genres — tend to generate long unbroken exchanges with tag-said-tag-said rhythm that reads flat over a full scene. If you write romance specifically, cross-reference this pass against Romance Beat Sheet: Where AI Drafts Usually Break, since pacing collapse in that genre usually clusters around the same beats — the black moment, the reunion — every single time.
For genre-specific mechanical tics, it helps to know what your genre's AI drafts tend to do wrong by default. Thriller with AI drafts often over-explain tension instead of trusting short sentences to carry it. Horror with AI drafts sometimes over-describe dread instead of withholding it. Knowing your genre's default failure mode means you can add a targeted line to your rung 2 prompt instead of relying on a generic mechanics check.
Rung 3: The line-polish pass — voice, word choice, rhythm
This is the last rung, and it's the only one where you actually want the AI touching individual sentences for style. By now the scene works structurally and reads cleanly at the mechanical level — so a line edit here won't accidentally erase a plot point or introduce a new pacing problem, because those were handled already.
The risk at this stage is genericization — AI smoothing your prose into a voice that sounds like everyone else's AI-polished prose. To prevent that, anchor the prompt to specific voice traits, ideally pulled straight from your story bible or a passage you know sounds like your book.
Structure and mechanics are both finalized for this chapter — don't change scene order, causality, or sentence patterns I've already fixed. Do a line-level polish only. My protagonist's internal narration voice is dry, clipped, and slightly self-deprecating — here's a paragraph from chapter 2 that nails it: [paste example]. Match that rhythm and word choice. Flag any word or phrase in this chapter that sounds more formal or more "AI-generated" than that sample, and suggest an alternative that fits her voice. Do not add new description or lengthen any sentence.
The anchor paragraph does most of the work in that prompt. Without it, the model defaults to a kind of generic literary smoothness — competent, but voiceless. With it, you're giving the AI a concrete target instead of an abstract instruction like "make the voice punchier," which means different things to a model than it does to you.
This is also the rung where you hunt filter words — "she felt," "he noticed," "it seemed" — that dilute close third person. And it's where you check that dialogue sounds like distinct people rather than one writer's voice split across quotation marks. If your manuscript has a large cast, running a dedicated character consistency check after the line-polish pass catches voice bleed the polish itself might introduce — sometimes making everyone sound better also makes everyone sound the same.
One caution: never run rung 3 on a chapter before rung 1 is done. I've seen writers get gorgeous, tightly voiced prose wrapped around a scene that doesn't need to exist. Polish is the most seductive rung because the output looks finished. Don't let "looks finished" fool you into skipping the rungs that check whether it actually is.
Building your rung checklist so passes don't collide
The ladder only works if each rung stays honest about its scope. The easiest way to enforce that is a written checklist you paste into every revision prompt, so the model is reminded of the boundary every single time — not just the first time you ask.
- Rung 1 checklist: scene want/outcome identified, causal link to next scene confirmed, stakes stated explicitly, filler scenes flagged, cross-chapter continuity checked. No prose comments allowed.
- Rung 2 checklist: sentence-structure repetition flagged, dialogue tag variety checked, paragraph length and pacing flagged, no scene-level or word-choice changes allowed.
- Rung 3 checklist: voice-sample anchor provided, filter words flagged, dialogue distinctiveness checked, no structural or mechanical changes allowed.
Keep this checklist in the same document as your story bible so it travels with the project. When you sit down to edit a book with AI, you're not reinventing the prompt structure each time — you're pulling the same three scoped prompts and swapping in the current chapter.
It's worth building this checklist once per genre, too, since the failure modes differ. A LitRPG writing draft needs a rung 1 check specifically for stat/progression consistency alongside scene function. An mystery novel with AI draft needs rung 1 to verify clue placement and fair-play logic, not just stakes. A science fiction with AI manuscript often needs a rung 1.5 check for worldbuilding consistency — pair that with fantasy worldbuilding tools or, for magic-based conflict resolution specifically, Fantasy Magic Systems: Constraints AI Will Respect, which covers a related but distinct problem: making sure your magic's rules don't get violated for plot convenience.
If you're drafting from scratch rather than revising an existing manuscript, the same rung logic should shape how you prompt the first draft too — the beginner handbook writing fiction with AI and the general write a book with AI guide both cover how to set expectations early so your rung 1 pass has less to fix later. And if you're deciding between tools for managing this kind of staged workflow, Entangled Text vs ChatGPT walks through why keeping revision context attached to your manuscript matters more than raw model quality.
Once all three rungs are done, don't stop there. A structurally sound, mechanically clean, well-voiced chapter still benefits from outside eyes — run it through a Beta Reader Workflow for AI-Assisted Manuscripts before you consider it locked. And if you're getting close to publishing, the ladder's final output should feed directly into your KDP Upload Checklist After an AI-Assisted First Draft or the broader Publish, Launch & Distribute Your Book: Complete Author Guide — since a clean rung 3 pass is usually the last stop before formatting for upload.
If you're revising an opening chapter specifically, run the rungs in the same order but weight rung 1 harder — Opening Hooks for AI Drafts: Fix the First Page Fast covers why first-page structural problems (a weak hook, delayed stakes) do more damage than anywhere else in the book, and why polishing prose on a page 1 that opens in the wrong place is wasted effort. Genre matters here too — a YA novel with AI opening needs voice to hit almost immediately, while historical fiction with AI can front-load more scene-setting before the hook lands. Track your revision time against a budget for AI drafting a full novel if you're running these passes at scale across dozens of chapters — three prompts per chapter adds up, and knowing your per-chapter cost keeps the ladder sustainable rather than a spiral of "just one more polish pass." For the technical side of running staged prompts against your manuscript, the Entangled Text documentation covers how to keep chapter context loaded across separate revision passes without re-uploading each time. For planning how long each chapter should run before you even reach revision, the Chapter Word Count Planner helps you catch pacing problems at the outline stage, before they ever reach rung 1.
Start with one chapter you already think is "done." Run rung 1 alone, and only rung 1 — no prose comments allowed. You'll likely find at least one scene that's beautifully written and structurally hollow. That's the ladder working exactly as intended: catching the expensive problem before you spend an hour polishing a sentence that shouldn't have survived the first rung.