What AI feedback on your manuscript can (and can’t) tell you
· The StoryProp team
There is a moment, usually late, when you have a full draft and no idea what you have. The pages are there. The shape is not visible from inside. This is the moment writers reach for a reader, and increasingly for a machine, and the first question is almost always the wrong one: is it good? Ask that and you will get an answer: fluent, encouraging, and worth very little. Three questions it does answer well, because all three are checkable against the page: where you repeat yourself, where the manuscript contradicts itself, and what shape your chapters make when you lay them end to end. Taste, boredom and originality are not on that list, and it will pronounce on all three anyway. The useful skill is knowing which questions a language model can actually answer about your manuscript, which ones it will answer badly while sounding certain, and how to hold both kinds of response at arm’s length long enough to decide what to do.
What machines are genuinely good at seeing
A model reading a long manuscript is doing pattern recognition at a scale and a patience level no human reader will match on a first pass. That is not nothing; it is most of what a structural read consists of. Repetition is the clearest case. You have a verbal tic and you cannot see it, because it sounds like your voice to you. Characters shrug, exhale, look away. Three chapters open with weather. The word “somehow” turns up on nearly every page of a book that should contain it twice. Ask for every instance and you get a list — not an opinion, a count, and counts are checkable.
Consistency is the second case. A long project accumulates small contradictions that no one holds in memory at once. The cup is chipped on the handle in chapter two and on the rim in chapter nineteen. Mara’s sister is mentioned in passing early and is an only-child fact later. Daniel drives to the house in one scene and has never learned to drive in another. These are not craft judgments. They are collisions between two statements, and a reader with total recall of the text will find them.
Pacing shape is the third, and the most interesting. Ask a model to summarize what happens in each chapter in one sentence, then read the list straight through. You will see immediately if six consecutive entries are all some version of “they talk about the house.” You will see the place where three chapters of plot are compressed into one, and the place where one scene has been spread across four. The model did not tell you your pacing was bad. It handed you a rendering of your book at a zoom level where the problem becomes visible.
Notice what these three have in common. In each case the machine supplies an observation that is verifiable against the page, and you supply the judgment about whether it matters. Repetition can be a motif. A contradiction can be an unreliable narrator. A slow stretch can be exactly the stretch the book needs. The note is data. The decision stays yours.
What it can’t tell you, and will happily tell you anyway
The category a model handles worst is the one writers most want handled: whether the book is worth reading, and what it is for. A manuscript has an argument, even a novel — some claim about people that the story exists to demonstrate or complicate. A model can often name that claim in the abstract, the way jacket copy does. What it cannot tell you is whether the book earns it, or whether the version on your pages is the interesting version or the obvious one. Taste is an assertion about value, and value is not recoverable from patterns in text.
It also cannot tell you about the reading experience, because it did not have one. Ask whether the middle drags and you will get a plausible answer, but the model did not get bored. Boredom is a fact about a person on a Tuesday evening. The same goes for surprise, and for staying up past midnight because you had to know. When a model reports that a scene is powerful, it is describing the features that scenes of that type usually have. That is a genre inference wearing the costume of a response.
The third gap is originality. A model is extremely good at recognizing what a thing resembles and extremely poor at recognizing what has never been done before. Its sense of the good sentence is downstream of the frequent sentence. Push a manuscript steadily toward what such a reader approves of and you push it toward the center of everything already written.
The failure mode to watch for is confident coverage: a note that answers the question you asked, in the register you asked it, with no signal that it fell outside the answerable range. Nothing marks which of the notes you just received are counts and which are guesses. Sorting them is your job, and it is essentially the whole job.
Ask narrow questions
The single change that most improves machine notes is replacing evaluation with description. “Is this chapter working” invites a verdict. “What does the reader know at the end of this chapter that they did not know at the start” invites a report you can check against the page.
Some questions that reliably produce usable answers: list every scene in which Mara makes a decision; show me where Daniel’s motivation is stated on the page, as opposed to implied; which chapters contain no dialogue; summarize the antagonist’s plan using only what is written, not what is suggested; trace every mention of the house’s ownership in order. Each of these is answerable from the text and falsifiable by you. And each has a way of answering the evaluative question sideways. If you ask where a motivation is stated and the honest answer is nowhere, you have learned something no amount of asking whether the character feels believable would have surfaced.
The technique generalizes. When you catch yourself wanting a grade, translate it into an inventory. Instead of asking whether the pacing is right, ask for a list of what changes in each chapter. Instead of asking whether the dialogue is good, ask which lines could be moved to a different character without anyone noticing. Instead of asking whether the ending satisfies, ask which promises made in the first fifty pages get answered, and where.
Triangulate, don’t obey
Machine notes are one instrument, and instruments are read against each other. A note from a model that a human reader later confirms — someone who stumbled in exactly that spot — is a real problem in the draft. A note that no human reader ever independently produces is a hypothesis you are welcome to discard. So split the questions by what each kind of reader can hold: ask people where they lost interest, where they stopped believing someone, where they put the pages down and did not pick them up again, and ask the machine for the things no person keeps in memory across four hundred pages.
Your own reading counts as an instrument too, and it has the most standing. You know what the book is for. When a note contradicts your intention, the useful move is neither to comply nor to dismiss, but to ask why the text produced that reading in the first place. Often the intention is intact and the execution leaked. Occasionally the note is simply wrong about a book it was never going to understand.
The blandness trap
The specific danger of revising toward machine approval is that every individual suggestion is defensible and the sum of them is a book with the strangeness sanded off. Smooth out the long sentence. Explain the ambiguous gesture. Cut the digression that serves no plot function. Each edit is locally reasonable. Enough of them and the manuscript reads like everything else.
A practical guard. Before a revision pass, write down the handful of things about this book that are not like other books — the voice, the structural risk, the character who refuses to be likable, the thread you intend to leave unresolved. Any note that touches something on that list needs a stronger reason than a reader suggested it. Any note that touches nothing on the list is cheap to try. The same applies to the machine’s own corrections: a tool that quietly rewrites an intentional oddity into an ordinary one has destroyed information, which is why it matters whether your working setup surfaces contradictions or silently resolves them.
It helps enormously if trying a note is reversible. A change you can undo cleanly is a change you can be brave with, so the mechanics are worth getting right: every draft kept, passes grouped so you can see what a single sitting changed, two versions set side by side, an earlier one restored without disturbing the work you did after it — the same machinery a revision pass runs on. What matters for feedback is the effect on judgment. When the cost of acting on a bad note is high, writers tend to either refuse every note or accept every note. Neither of those is discernment.
You remain the editor of your editors
The frame worth holding is that a model is not a beta reader and not a critique partner. It is a very fast, very literal reader with total recall, no taste, and no stake in the outcome. That combination is genuinely valuable for one class of problem and actively misleading outside it. So use it where the answer is checkable. Let it find the repetitions, the contradictions, the chapters where nothing changes, the promise you made on page nine and forgot by page two hundred. Let it build the inventories that make your own structural sense operational. Then close the notes and decide — because the question of what this book is for was never available to any reader but you.
Notes are only as good as the memory behind them, and only as safe as your ability to walk a change back. StoryProp keeps the confirmed canon, the decisions you’ve made, and every earlier pass on file, so when you act on a note the change lands where you pointed it and the rest of the manuscript stays put.