The Right Way to Get Feedback From AI on Your Manuscript

Paste a chapter into most AI tools and ask "what do you think?" and you'll get a warm bath of praise. "This is a compelling opening with vivid imagery and strong emotional stakes." It feels good for about four seconds, then you realize it told you nothing you can act on. Worse, it told you the same thing it would have told a genuinely broken chapter.

The default behavior of these models is to flatter you. They're tuned to be agreeable, and agreeable is useless when you're trying to find what's wrong with your own book. Getting real AI feedback on writing is a skill, and it comes down to one thing: refusing to let the model be nice.

Why the praise is worthless

Language models are trained to be helpful and pleasant, which in practice means they default to encouragement. Ask an open question and you get a hedge. The model doesn't want to risk being wrong about your intent, so it stays vague and positive.

That instinct is exactly backwards for revision. A good critique partner isn't there to make you feel good. They're there to tell you the dialogue on page three sounds like two robots reading a contract. If you want that, you have to force it, because the model will not volunteer it.

So stop asking "is this good?" That question has no useful answer. Start asking questions that can only be answered with something specific and concrete.

Force specificity or get nothing

The single biggest upgrade to your prompts is constraint. Vague questions produce vague answers. Narrow the target and the feedback sharpens immediately.

Compare these two:

What do you think of this scene?

In this scene, mark the three weakest sentences and explain what makes each one weak. Then point to the exact line where the tension peaks, and the exact line where it goes slack.

The first gets you a book report. The second gets you something you can fix tonight. You've given the model a job with edges, and you've forced it to commit to specific lines instead of floating above the text.

This is the same instinct a good AI writing coach trains in you: interrogate the work, don't admire it. A few prompts that reliably produce real notes:

• "Quote the first sentence where you got bored. Don't soften it."
• "Where did I tell the reader something I should have shown? Give line numbers."
• "Which character's dialogue could I swap with another character's and not notice? That means their voices aren't distinct."
• "What question is the reader asking on page one, and is it still pulling them by page three?"

Every one of those demands evidence. The model can't wriggle out with "great pacing!" because you've asked it to point at the text.

Make it play a skeptical editor

The fastest way to kill the flattery reflex is to assign a role with an attitude. Models follow personas well, and a hostile one cuts straight through the niceness.

You are a tired, skeptical acquisitions editor who has read 200 manuscripts this month and rejected 195 of them. You are looking for reasons to stop reading. Read this opening and tell me exactly where you'd put it down, and why.

The tone shifts completely. Now it's hunting for weaknesses because that's the job you gave it. You can run variations: a genre reader who's seen every trope, a line editor who hates adverbs, a reader who skims. Each lens surfaces different problems.

One caveat. A skeptical persona will sometimes invent flaws to satisfy the role, the same way a flattering one invents virtues. Treat its complaints as leads to investigate, not verdicts. The point isn't to obey the editor. It's to get a stream of specific objections you can test against your own judgment.

Separate line-level notes from structural notes

Here's a mistake that quietly wrecks revisions: asking for everything at once. When you say "give me feedback," the model mixes a comma splice on page two with the fact that your protagonist has no goal, and treats them as equally important. They are not.

Run two distinct passes, and never blend them.

Structural pass first. Zoom out. Does the scene have a point? Does something change between the first line and the last? Where does the chapter sag? Fixing structure can delete whole paragraphs, so polishing sentences before you've settled the shape is wasted effort. Ask: "Ignore the prose entirely. Tell me what this scene accomplishes for the plot and what changes by the end. If nothing changes, say so."

Line pass second, and only once the bones are set. Now you want the granular stuff: weak verbs, repetition, rhythm, clichés. "Forget the plot. Find every sentence that relies on a weak verb plus an adverb where one strong verb would do. List them."

Doing structure before line means you don't lovingly polish a scene you're about to cut. This ordering matters enough that it deserves its own habit, and it's the backbone of revising with AI without flattening your style.

Push back, always

Whatever the model says, your first move is to challenge it. This is non-negotiable, and it's where most people stop too early.

The model is confident and wrong constantly. It will tell you a scene drags when it's actually doing slow-burn tension on purpose. It will flag a line as "confusing" when it's deliberately withholding. If you accept every note, you'll sand your book down to the model's idea of competent, which is the statistical middle of everything it has read.

So argue:

You said the opening is slow. I made it slow deliberately to build dread before the reveal on page four. Does it still work for that purpose, or is it genuinely dead weight? Be honest.

Sometimes it backs down and admits your version is stronger. Sometimes it holds the line and gives a better reason, and you realize it was right. Either outcome is useful, because you've stress-tested the note instead of swallowing it. That back-and-forth, you defending a choice and the model probing it, is where the actual thinking happens. It's also why an AI tool that holds your whole manuscript as context, like what Polyz does, beats pasting fragments into a generic chatbot: it can see that the "slow" opening pays off in chapter two.

A repeatable feedback loop

Put it together and you have a routine you can run on any chapter:

1. Structural pass. Ask what changes in the scene and where it sags. Ignore prose.
2. Skeptical-editor pass. Assign a hostile persona and ask exactly where they'd stop reading.
3. Line pass. Hunt weak verbs, repetition, and clichés with line numbers.
4. Push back on every note that touches a deliberate choice. Make the model defend itself.
5. Decide. You keep the notes that survive your judgment and ignore the rest.

The model never gets a vote on the final draft. It generates pressure, finds blind spots, and tells you things your too-close eyes can't see anymore. You make every call. That division of labor is the whole game: the feedback gets you to a better book, but the taste that picks which feedback to use stays yours.

Get that frame right and AI becomes the most patient, available reader you'll ever have. Get it wrong and it becomes a yes-man with a thesaurus.

---

Polyz reads your whole manuscript, gives you specific craft feedback, and pairs it with an AI writing coach that helps you get better instead of just handing you fixes. Start a free trial.