Writing with an agent

AI for outlining vs. AI for drafting: two different jobs

August 16, 2026 · The StoryProp team

Writers ask two things of an AI tool, usually in the same breath: help me figure out my story, and help me get it onto the page. The requests sound alike. They point in opposite directions. Figuring out wants the space of possibilities held open — more options, better questions, structure made visible. Getting it onto the page wants that space closed — one scene, in your register, honoring every decision you have already made. In the language of design, outlining is divergent and drafting is convergent, and a tool tuned for one will, left to itself, quietly sabotage the other. Scale settles one part of this for you: Anthropic publishes a 128,000-token ceiling on what Claude Opus 5 can emit in a single response, which at Anthropic's own stated ratio of 555,000 words per million tokens comes to roughly 71,000 words — so a novel arrives in many passes whatever tool you use. What makes each of those passes convergent is not the ceiling. It is that every one of them has to honor the decisions the last one made.

Outlining is divergent work

When you are outlining, the last thing you need is an answer. You need a spread. Suppose you are circling a quiet domestic novel — a woman, call her Mara, returning to her childhood house after her brother Daniel stops answering the phone. The outline question is not what happens in chapter three. It is: how many different books could this be? The one where Daniel has been in the house the whole time. The one where he has been gone a year and everyone knew but her. The one where the house itself is the problem. An AI tool to outline a story earns its keep by multiplying these possibilities honestly — not by picking one for you.

The second thing outlining wants is questions you had not thought to ask. What does Mara actually want from the visit, and what would she say she wants? Whose version of the family story is the book going to endorse? What must be true by the final chapter for the ending to land? A good outlining session with an AI reads less like dictation and more like being interviewed by an unusually patient editor. The value is not in the machine's answers. It is in what the questions force you to decide.

How an outline becomes a record of commitments rather than a list of events is its own craft, and it has its own piece: outlining as decision-recording. The narrower point here is that at this stage, the machine's job is to make sure you are choosing rather than defaulting. Divergence is a service. At the wrong moment, it is also a liability.

Drafting is convergent work

At drafting time, invention is exactly what you do not want. You have decided: chapter three, Mara alone in the kitchen, washing the chipped cup that was Daniel's, realizing she is rehearsing an argument with someone who is not there. The job now is sentences — sentences that sound like the previous forty pages, that know the cup is chipped and why it matters, that do not introduce a helpful neighbor or a sudden memory of the mother because the scene struck the model as a little quiet.

This is what people are really asking for when they go looking for an AI chapter writer: not a machine with ideas, but a machine with discipline. Convergent drafting is compliance work — with canon, with tone, with what the reader knows at this point and no more, with the rhythm of your sentences rather than the median rhythm of published English. It is harder than it looks, and it is a different hardness than brainstorming.

Convergence also implies humility about scope. A drafting pass should produce the scene you asked for, at roughly the length you asked for, and stop. And when an instruction collides with something established — Daniel's cup was blue in chapter one, green in today's note — the useful behavior is to surface the conflict, not to quietly pick a color and move on. In drafting mode, an unrequested decision is a small betrayal.

Why a tool great at one job flubs the other

The failure modes are mirror images. Point a divergence-tuned tool at a drafting task and you get invention smuggled into prose: new objects, new gestures, a subplot arriving like an uninvited guest, every paragraph offering options in the costume of a scene. This is why an AI outline generator for novels, asked for chapter three, so often returns something that reads like a pitch — energetic, plausible, and not your book.

Point a convergence-tuned tool at an outlining task and you get the opposite disappointment: the median plot. Ask what could happen and it tells you what usually happens — estrangement, secret, confrontation, rain on the drive home. Systems built to complete patterns reach first for the most completable pattern, which is the most familiar one. At the moment you most need the strange option, you are handed the safe one, beautifully formatted.

General-purpose assistants are built to hold both jobs in one place, and that place has a documented shape. A chat is one undifferentiated scroll, and Anthropic's platform documentation names what happens as that scroll grows — “context rot,” its term for the way accuracy and recall degrade with token count, so that usable capacity is always smaller than nominal capacity. Its compaction mechanism, currently in beta, fires by default at 150,000 input tokens: the API writes a summary and then drops every content block before it from subsequent requests. Anthropic's stated rationale is not that space runs out but that response quality degrades as context grows. Inside the consumer Claude.ai product the same pressure is handled by summarising earlier messages to make room for new ones. Your chapter-one decisions do not vanish — they become a paraphrase, which is exactly the form in which a specific detail (Daniel's cup, chipped, blue) stops being binding.

It also helps to know how much room you actually have, because the number in the headlines is the developer figure, not the one behind your chat window. OpenAI publishes a 1,050,000-token context window for GPT-5.6 Sol on its API and 272,000 tokens for the same model on the ChatGPT Business plan — roughly a quarter. Anthropic's gap is narrower and turns on which model you are using: paid Claude.ai plans get the full 1,000,000 tokens for Claude Opus 5 and Sonnet 5, the same window those models carry on the API, but 500,000 for Opus 4.8, whose API window is also 1,000,000, and 200,000 for its remaining models, which it describes as about 500 pages of text or more. Those are generous numbers, and a novel-in-progress plus its notes will still walk into them by the middle of a project.

The fuller sorting of tools — generators, assistants, partners, and what each is actually for — is a taxonomy of its own. The question worth asking in the moment is simpler: which job does this tool believe it is doing right now, and do you agree?

Outlining help that doesn't take the wheel

A version of the same worry arrives as a search query: what AI can help me outline a book without writing it for me? Behind it is a legitimate fear — that accepting help with structure means surrendering the story, that every suggestion is a lever pulling the book toward the machine's taste. The answer turns on a single property: does a suggestion become part of the book just by appearing on screen?

In an ordinary chat, it effectively does. Everything scrolls into the same undifferentiated history — your decisions, the machine's musings, the options you waved away — and whatever gets generated next is shaped by all of it equally. The consumer assistants have begun to separate the two, and their own descriptions are instructive. OpenAI documents two distinct mechanisms in ChatGPT: saved memories, the details you explicitly asked it to remember, kept until you delete them; and referenced chat history, which is inferred from past conversations and which OpenAI says can change over time as ChatGPT updates what is more helpful to remember. OpenAI's published advice is to use saved memories for anything that must always be remembered. That is the difference between a decision and a suggestion, stated by the vendor. Anthropic scopes it differently again: Claude's memory is written as individual entries organised into categories, with each Project holding its own separate memory space.

StoryProp is built on that boundary rather than layered over it. What you agree to gets written down; what you don't stays a suggestion. Proposals sit apart from canon — characters, story shape, confirmed style rules — until you promote them, and the agent reads the promoted records before it writes a line. Research is saved with its sources and dates, in its own place, so something you looked up on Tuesday never quietly becomes a fact about your world on Friday.

Delegating prose labor, in other words, is not delegating authority. You own the outline; the decisions in it are yours whether you typed them or nodded at them. What you hand over is the typing — the long, patient conversion of decisions into sentences — and even there the register stays yours, because decisions about register can be recorded like any other. A writer who keeps this straight can accept a great deal of help and lose nothing that matters.

The handoff between modes

The moment that decides whether the whole arrangement works is the handoff — the point where you stop asking what could happen and start saying write what happens. A scene is ready to cross that line when you can state its job in one sentence, name what changes between its first line and its last, and say what the reader knows afterward that they did not know before. If you cannot do those three things, you are not ready to converge, and no drafting tool will rescue you. It will simply invent the missing decisions and present them to you as finished prose.

The handoff is not a one-way door, and it happens scene by scene, not once per book. You will converge cleanly on chapter three and be thrown back into divergence at chapter fourteen, when the plan meets the book you have actually written and loses. This is normal — it is most of what writing a novel is. The Science Fiction and Fantasy Writers Association puts the floor for a novel at 40,000 words in its Nebula rules, and at any plausible scene length that is dozens of separate handoffs, each with its own chance to re-open. What you want is a workflow that makes re-opening cheap: draft in small pieces, keep every pass, and when a structural question resurfaces, duplicate the document or the whole work and let two answers compete on the page instead of in your head.

Those are the mechanics that make toggling practical rather than theoretical, and they are what StoryProp's feature set is built from: recorded decisions the drafting pass actually consults, every draft kept as sessions and passes, side-by-side comparison between versions, non-destructive restore and whole-pass undo, and duplication of a document or a whole work when you want two answers to a structural question to compete in full. That is the difference between a handoff and a hostage exchange.

Two jobs, one book. The writers getting the most out of these tools are not the ones with the cleverest prompts. They are the ones who always know which job they are asking for — who let the machine run wide while the outline is wet, and hold it strict once the outline has set. Keep the modes distinct and the help compounds: the outline stays yours because you chose every line of it, and the draft sounds like you because the machine was told, in writing, what you sound like. That is a very good place to be writing from.

The cleanest version of this workflow is one where the boundary is built in: suggestions stay suggestions until you promote them, and every drafting pass is written from the decisions you actually recorded. That is the working agreement StoryProp keeps — you hold the outline, it handles the typing.

← All posts