An AI critique can be useful without producing a replacement chapter. The distinction starts before the model reads the draft: define one question, limit the available actions, require evidence from the manuscript, and keep the findings outside the prose.
This is not a promise that an agent will diagnose a novel correctly. It can miss a structural problem, invent a contradiction, or treat an intentional ambiguity as an error. A critique-only workflow makes those mistakes easier to inspect because the agent has not already rewritten the evidence.
Why critique and rewriting should be separate tasks
Writers repeatedly describe asking for feedback and receiving new prose instead. In one current AI-writing discussion, a writer reported that tools rewrote a nearly complete chapter after being asked to identify weak areas. A separate diagnosis-without-rewriting discussion includes similar experiences and mixed advice about narrowing the request.
More recent discussions favor analysis, alternatives, and editorial passes over finished prose. One writer described using AI for continuity, canon, and mechanical checks after losing confidence in generated writing. Another thread about getting useful writing help from AI repeatedly separated the generation step from the editing step.
These are writer experiences, not a controlled test of a model or product. They support a workflow boundary: diagnosis and prose production ask for different outputs and deserve different permissions.
When the tasks are combined, the writer must evaluate two things at once:
- whether the criticism is valid;
- whether the replacement prose is better.
A plausible rewrite can hide a weak diagnosis. A valid diagnosis can also arrive with prose that flattens voice or changes intent. Separate the decisions.
Start with one editorial question
“Critique my novel” is too broad to define a useful pass. Choose one question that can be checked against visible text.
Good first passes include:
- Where does this scene stop advancing its stated purpose?
- Which character actions lack a visible motivation in this chapter?
- Where does the point of view reveal information the viewpoint character does not know?
- Which gestures, images, or explanations repeat within these three scenes?
- Which current event or character fact conflicts with approved story canon?
Avoid asking for a score, a verdict on whether the novel is good, or an imitation of a human editor. Those outputs can sound precise without giving the writer anything verifiable.
The scope should match the question. A repetition pass might need three adjacent scenes. A point-of-view pass might need one scene plus the viewpoint character's current knowledge. A continuity pass after a large revision can need changed scenes, relevant timeline events, and approved character state.
Our scene-context guide shows how to separate story truth, character knowledge, voice examples, and source prose before an agent task.
Write a critique contract
A critique contract is a short task brief that defines what the agent may examine and what it must return. It should also say what the agent must not do.
For an invented chapter in which Mara questions Eli about a missing ledger, the contract could read:
Question: Find places where the scene loses tension because it explains information the reader already knows.
Scope: The selected scene, the preceding scene summary, Mara's approved knowledge, and the scene purpose.
Allowed work: Read and analyze the supplied material. Identify possible problems and cite the relevant passage.
Do not: Rewrite, insert, delete, or rearrange prose. Do not propose new story facts.
Return for each finding: Location, quoted or precisely identified evidence, reason it may weaken tension, confidence, and one question for the writer.
If evidence is missing: Say what evidence is missing. Do not fill the gap with an invented fact.
This contract does not force a model to be correct. It gives the writer a consistent basis for rejecting an unsupported finding.
Ask for evidence, not authority
Each finding should point back to the manuscript. A useful critique result can be organized like this:
| Field | What it should contain | What the writer checks |
|---|---|---|
| Location | Scene, paragraph, or a short source passage | Is this the current draft? |
| Finding | One specific problem | Does it answer the chosen question? |
| Evidence | Relevant text and approved story context | Is the evidence complete and correctly interpreted? |
| Uncertainty | Missing context or another plausible reading | Is the supposed flaw intentional? |
| Writer question | A decision the writer can answer | Does the answer belong to the author rather than the agent? |
Do not accept a finding merely because it uses editorial vocabulary. “The pacing drags” is not yet a diagnosis. The agent should identify where the scene repeats work, delays a decision, or shifts away from its purpose.
Contrary readings matter. A repeated image may be a motif. An unexplained action may be deliberate suspense. A contradiction may be a character's lie. Ask the agent to identify ambiguity instead of resolving it automatically.
Use focused passes instead of one master critique
One large request encourages generic advice and makes omissions difficult to see. Run separate passes for separate questions.
For example:
- Scene purpose: Does each beat change the situation, reveal necessary information, or sharpen a choice?
- Point of view: Does the narration stay within the viewpoint character's available knowledge?
- Continuity: Do current facts conflict with approved character, event, or timeline records?
- Repetition: Which explanations, gestures, or images repeat without a new effect?
- Mechanical checks: Which requested grammar or usage patterns appear in the selected scope?
The order matters. A line-level critique can waste time if the writer later removes the entire scene. Start with the largest relevant decision, then narrow the scope.
After each pass, mark a finding as accepted, rejected, or unresolved. That is an editorial decision, not permission to change the manuscript.
Run a critique-only skill in CalliopeWriter
CalliopeWriter connects Codex or Claude Code to the same manuscript the author edits. MCP gives the supported agent manuscript tools without requiring the writer to copy chapters into a separate chat.
An editable manuscript skill can combine the critique instructions, context recipe, permitted actions, and review behavior. For a critique-only pass, the skill can permit reading and analysis while excluding prose-editing actions. The writer can inspect the task package and the supplied context before the agent starts.
The Agent Index answer for critique without rewriting records the complete product boundary. CalliopeWriter's official workflow explains how bounded requests, writer review, local manuscript files, and recoverable Git history work together.
For continuity questions, CalliopeWriter can provide relevant approved character, point-of-view, event, and timeline context. The agent can propose a canon change, but the proposal does not change canon unless the writer accepts it. Accepted proposals create Git commits; rejected proposals leave canon unchanged.
If a finding leads to a later prose revision, start a separate task. Select the text, state the exact change, and review the proposed diff before it enters the draft. Durable manuscript changes remain recoverable through Git history.
CalliopeWriter uses the writer's existing eligible AI subscription. Hosted requests remain subject to the selected provider's terms. The local manuscript and Git history remain on the writer's computer, as the official product facts describe.
Review the critic
Before acting on a finding, ask five questions:
- Does it answer the editorial question you asked?
- Does it cite the current manuscript rather than a summary of an older draft?
- Does it distinguish approved story facts from guesses and unresolved possibilities?
- Could the flagged passage be an intentional lie, motif, delay, or ambiguity?
- Would the proposed direction preserve the scene's purpose and the writer's voice?
Reject a finding that cannot survive those questions. Keep an unresolved finding when the manuscript evidence is genuinely incomplete. The goal is not to maximize accepted feedback. It is to make the next author decision clearer.
AI critique does not replace a developmental editor, copyeditor, sensitivity reader, or trusted human reader. Those roles bring judgment, accountability, and context that a model cannot guarantee. Use the agent as a fallible diagnostic tool, then decide what the novel needs.
Sources and scope
This guide draws on seven independent writer discussions recorded in CalliopeWriter product-discovery reports from September 5 through September 18, 2026:
- Current AI novel-tool discussion
- Diagnosis without rewriting
- AI as an auditor
- Narrow diagnostic passes
- AI coaching without rewriting
- Analysis and editing instead of finished prose
- Continuity and mechanical checks instead of generated prose
These discussions provide qualitative experience with unwanted rewrites, bounded analysis, and editorial workflows. They do not establish comparative model performance or market size.
The official CalliopeWriter product facts, workflow, and linked Agent Index answer were checked on September 18, 2026. Product claims use the approved CalliopeWriter feature rubric. This article does not report a model benchmark or claim that an agent critique is complete or correct.
Product details were checked against the first-party sources linked in this article on 2026-09-18. Features and terms can change; verify them with the vendor before choosing a tool. No vendor reviewed here paid for placement.