Writing with an agent

How to write a novel with AI, without losing your voice

August 16, 2026 · The StoryProp team

The bad version of writing a novel with AI is well documented: ask for chapters, receive competent mush, lose interest around the fourth one because nothing in it is yours. The failure isn't the technology. It's the method — treating a drafting partner like a vending machine. Here's the method that works, and how StoryProp is built around it.

Start smaller than feels productive

Don't ask for a chapter. Be exact about what that request is: Reedsy gives the working convention for an adult-fiction chapter as 2,000 to 4,000 words, and The Manuscript Academy's 2026 guide to what agents want, sourced to an agent at Transatlantic Literary Agency, puts commercial fiction at 80,000 to 100,000 words. So a chapter is a couple of thousand words you have to read, judge and mostly rewrite before you learn anything, and drafting a book that way gives you twenty to forty moments in the whole project where you get to steer.

So ask for the opening situation instead — a paragraph or two. Small drafts keep you in the director's chair: you can steer after every increment, and your reactions to concrete sentences teach the tool your taste far faster than abstract instructions do. In StoryProp this is the default rhythm on purpose; a full chapter is something you request in so many words, not something that happens to you.

Lock decisions one at a time

The single highest-leverage habit: when something is settled, say so, and make sure it's recorded somewhere more durable than the scroll-back. Premise, a character's secret, the rule your ending must keep. Equally important, keep everything else explicitly unsettled — a wishlist, not a checklist. You don't need the spine, the midpoint or the ending to start. You need one locked premise and an opening worth reacting to. And recorded should mean somewhere the tool actually consults before drafting — a transcript treats your accepted ending and your rejected one identically, which is a large part of why chatbots lose the plot over a long project.

Revise by pointing, not regenerating

Never rewrite a scene wholesale when one exchange is the problem. Point at the passage — the sentence, the beat, the line of dialogue — and say what you want changed. Targeted revision preserves everything you've already made true and confines the machine's contribution to the thing you actually asked for. If your tool can't leave the rest of the page alone, that's a tool problem, and it's the one to solve first. StoryProp's revision is pointed by design: it changes what you point at and leaves the rest alone, and where a change would contradict something you've already settled, it surfaces the contradiction rather than quietly fixing it.

Here's what that looks like in practice. Suppose the draft reads: “Mara set the chipped cup on the counter. ‘You knew about the lease,’ she said. ‘You knew for months and you let me sign anyway.’ Daniel looked at the floor and said he was sorry.” The scene works until the last beat — Daniel folding is wrong for a man who spends the whole book unable to apologize. So you select that one sentence and say: he doesn't apologize; he changes the subject to the cup.

The revision comes back: “Daniel picked up the cup and turned it so the chip faced away from her. ‘This was my mother's,’ he said. ‘You've been drinking out of it for a year and you never asked.’” Mara's lines are untouched, the counter is still the counter, and the only thing that moved is the thing you pointed at — a deflection now doing the character work the apology was undoing. That's the whole discipline of pointed revision: the smaller your request, the more of the page stays yours, and the easier it is to judge whether the change earned its place.

Sooner or later the draft will hand you something better than what you intended. You meant for Daniel to confess at the funeral; instead he almost confesses and loses his nerve, and you know — the way you know when an editor is right — that the failed confession is the stronger scene. The two reflexive responses are both wrong: rejecting the idea because it wasn't yours, or accepting it passively and drifting into a book you're no longer steering.

The useful test is whether you can say why it's better in the terms of the book you're already writing. “The failed confession is better because Daniel's whole pattern is almost-saying things” is taste at work: the suggestion revealed something your plan hadn't caught up to, and adopting it is authorship, not surrender. If the best you can manage is “it sounds impressive,” be suspicious; that's how manuscripts fill up with borrowed moves. Ownership was never about generating every idea yourself — it's knowing which ideas belong. So when one does, lock it, and next month it's a decision rather than a lucky accident.

Make the machine hold continuity, you hold taste

Continuity is bookkeeping, and bookkeeping is what machines are for — it's the center of what novel-writing software ought to take off your hands. Who knows what, as of which scene. What the house looks like. What got burned, buried or said. Delegate that ledger entirely — and insist on being flagged, not corrected, when a new passage contradicts it, because sometimes the contradiction is the revision you meant to make.

The same goes for facts. When the draft needs a real one — a poison's onset, a 1962 train timetable, how probate actually works — the research should happen without leaving the page, and it should come back with sources attached rather than as confident unsourced assertion. Keep researched material clearly separate from your story's canon. One is evidence; the other is what you decided.

Taste is the opposite: never delegated. Voice stays yours by feeding the record with confirmations rather than letting the tool infer on its own — praise the passage that sounds like you, correct the one that doesn't, and let confirmed preferences become durable rules while everything else stays observation.

Keep a rhythm across the weeks

A novel is written in sessions, and the space between sessions is where most methods quietly fail. The enemy isn't lost time; it's lost state — sitting down after two weeks away and spending the first half hour reconstructing what you'd decided, which is demoralizing enough that many people don't sit down at all. So design for re-entry: end each session by settling one small thing out loud, even if it's only “the next scene is Mara at the bank, and she doesn't get the loan.” That sentence is next week's on-ramp. Then let the project hold the rest — if your decisions, canon and drafts live in records the tool consults, re-entry is reading a short list, not re-reading a hundred pages.

The most precisely documented working novelist we have is Anthony Trollope, who kept a watch on the desk and published the totals in An Autobiography: 250 words a quarter hour, three hours each morning from half past five, and a self-imposed quota averaging forty manuscript pages a week — about 10,000 words — around a full-time job at the Post Office. The variance is the encouraging part. His recorded weeks ran from 5,000 words to 28,000. What held the total together was not even output but a fixed shape, including the half hour at the start of every session that he spent rereading yesterday's work before writing a new word — a re-entry cost paid daily, by a professional, with no gap at all. Sessions don't need to be long or regular to add up; they need to start where the last one ended, and that's a property of the system, not of your discipline.

Finish clean

The end of the process is a manuscript with your name on it — a professional file you'd send without explanation. No watermarks, no AI notices, no branded filenames. The reader's judgment should land on the book, and the byline on it should be exactly the one you chose. None of this hands the book to someone else. It makes the writing yours while removing the parts that were never writing: the bookkeeping, the re-explaining, the fear of breaking what already works. That is the promise worth demanding from a writing partner, and the one StoryProp is built to keep.

Small drafts, decisions locked one at a time, revision that touches only what you point at — that rhythm is the shape StoryProp takes: a permanent project conversation, what you agree to promoted to records the agent reads before it writes, and every draft kept, so a risky pass can be compared side by side and undone whole. Bring the smallest scene you have and direct it there.

← All posts