AI can comment intelligently on a scene it can see. The harder problem is judging a whole novel when the opening, midpoint, and ending do not fit comfortably into one working context. If you paste a large manuscript into a single prompt and ask for “a detailed critique,” the model may compress the middle, confuse setup with payoff, or produce advice that sounds editorial without being anchored to the book.
The solution is not a more theatrical prompt. It is a better workflow.
Why one giant prompt fails
A novel contains several kinds of context at once: who knows what, when a promise is introduced, how a relationship changes, whether a timeline closes, and what the reader is meant to expect. A chapter can appear slow in isolation but be exactly the pause the book needs after a major reversal. A detail that looks inconsistent in chapter four may be deliberately corrected in chapter twelve.
When a model has to read, remember, diagnose, quote, prioritize, and summarize a full book in one pass, those tasks compete. Even if the file is accepted, acceptance is not proof that every chapter received equal attention.
Step 1: create a stable chapter inventory
Start by extracting the text without changing the original. Assign every chapter and chunk a stable ID such as ch007-p002. Keep the IDs even if the analysis is interrupted. They become the address system for reports, quotes, and progress.
A good inventory records the chapter title, order, approximate size, POV, time and place, major event, and what changes by the end. At this stage, resist the urge to diagnose. You are establishing coverage.
Step 2: build the manuscript map
The manuscript map is the shared memory of the review. It should be concise enough to carry into every chapter analysis but specific enough to prevent generic feedback. Include:
- the premise, genre, audience, and reader promise;
- character goals, motivations, conflicts, and arc states;
- the main plot and subplot chains;
- chronology, locations, rules, terms, and knowledge boundaries;
- setups, promises, open questions, and later payoffs.
Update the map as later chapters clarify earlier ambiguity. Do not label an open question a plot hole until you know the book does not answer it.
Step 3: analyze chapters sequentially
Now review each chunk against both its local scene and the global map. Ask what the chapter is trying to accomplish, what changes, and what consequence moves forward. Then examine plot causality, motivation, scene tension, pacing, dialogue, POV, style, and reader orientation.
A useful finding has four parts:
- Diagnosis: the precise problem, not a vague label.
- Evidence: an exact short quote and stable location.
- Consequence: what the reader misunderstands, stops believing, or stops feeling.
- Revision direction: what needs to change without ghostwriting the author's sentence.
Save each chapter report immediately. A saved state file should distinguish complete and pending chunks so the work can resume after an interruption.
Step 4: look for contradictions between chapters
Cross-chapter review is where a full-manuscript workflow earns its keep. Compare the map and reports for:
- a character acting on information they have not learned;
- motivation that changes without a triggering event;
- setups that vanish or payoffs that were never prepared;
- travel time, age, weather, injury, or object continuity errors;
- POV distance and narrative rules that drift;
- a subplot whose scenes repeat the same beat instead of escalating.
Group repeated symptoms under one root cause. If five middle chapters drag because the protagonist has no active objective, the report should not pretend there are five unrelated pacing problems.
Step 5: verify every important claim
Models can miscopy or reconstruct a quote. Treat every quotation as data that must be checked against the prepared source. A practical evidence ledger records a finding ID, chunk ID, line range, exact quote, and severity. A validator can then confirm that the location exists and the quote matches.
Step 6: build the revision plan in dependency order
The final report should not be a chronological dump of chapter comments. Prioritize work so one revision does not invalidate the next:
- premise, structure, missing causality, and ending;
- character motivation and arc changes;
- scene order, pacing, and subplot integration;
- dialogue, exposition, POV, and line-level patterns;
- copyediting and proofreading after the story is stable.
A short example from a fictional book
In the synthetic three-chapter story The Bell Below, Mara turns a mysterious gear after thinking of her missing father. The action is dramatic, but the earlier chapters provide only second-hand facts about him.
Diagnosis: the final choice invokes a relationship that has not gained enough scene weight. Seed a concrete memory, object, or costly belief earlier so the same action reads as Mara's choice rather than the plot selecting the dramatic lever.
Notice what the finding does not do. It does not write the memory for the author. It identifies the missing narrative function and points to the dependency: seed first, then revisit the climax.
These questions also come from real publishing frustrations. Author Nadi Petrova describes how her experience preparing two books led to EditRoast. Her story is a useful companion to the technical steps above.
Use the free Manuscript Roast Skill
The free EditRoast Manuscript Roast Skill packages this process for Codex and Claude Code. It supports TXT, Markdown, and DOCX; saves the map and progress locally; checks full coverage; and validates quoted evidence. There is no registration or payment for the download.
If you want a quick browser-based first look before a full critique, use the free Manuscript Analyzer. For the next stage after developmental revision, read how to proofread an entire book with AI without losing your voice.