Writing with an agent

AI writing tools for beginners: how to actually start

August 16, 2026 · The StoryProp team

Most people open an AI writing tool for the first time, type something ambitious, and get back a wall of prose that is fluent, competent, and completely unlike anything they wanted to write. The natural conclusion is that the tool is wrong for fiction. Usually the tool is fine and the first session was wrong. You asked a stranger to write your book before you had told it anything about it.

There is a method for the first hour that works better, and it inverts almost everything the interface seems to invite. You start smaller than feels productive. You spend most of your typing reacting rather than instructing. And you decide, early and out loud, what your prose sounds like, before enough generated material accumulates to quietly answer that question for you.

It helps to name what you are sitting in front of, because the beginner's landscape is short. There are the general assistants — ChatGPT, Claude, Gemini — chat products that will draft fiction on request, and there are project-aware writing agents built around a manuscript, StoryProp among them. The difference that decides your first week is not fluency. It is what survives the conversation. Anthropic's help documentation says Claude summarises earlier messages to make room as a chat grows; its developer documentation describes a compaction step that fires by default at 150,000 input tokens, writes a summary, and drops everything before it from later requests. Anthropic also documents context rot: as the token count climbs, accuracy and recall degrade. Which means your voice rule from message three is a paraphrase of itself by message three hundred, unless it was written down somewhere other than the chat.

Start smaller than feels productive

The instinct is to test the tool at scale. Ask for a chapter. See what it can do. This is a reasonable instinct for evaluating a machine and a terrible one for starting a manuscript, because a chapter contains hundreds of small decisions you have not made yet, and the draft will make every one of them for you. By the time you read it, the decisions are load-bearing. Untangling them costs more than writing the chapter would have.

Ask for a paragraph instead. Or two lines of dialogue. Or a description of one room. The unit should be small enough that you can hold the entire thing in your head and say precisely what is wrong with it. That is the actual skill you are building in the first session, and it is much easier to build on eighty words than on eight hundred.

Here is a concrete first move. Take a scene you already half-know: two people, one room, one thing that needs to be said and is not being said. Describe it in three sentences of plain direction. Mara is washing dishes. Daniel comes in to tell her he has sold the house. He does not tell her. Ask for the first four lines only. Read them. Then say what is off.

What comes back will be too explanatory, almost always. The AI will have Daniel hover in the doorway and think about how hard this is. You say: cut the interiority, I want it in the hands and the objects. What comes back next will be better, and more to the point, you will have learned that this is a lever you can pull. That single exchange is worth more than a generated chapter.

React instead of prompt-engineering

Beginners burn enormous effort trying to write the perfect instruction up front — the paragraph of setup that will make the output come back right the first time. It rarely works, and it trains a bad habit, because the specification you need does not exist in your head yet. You discover it by seeing a version that is wrong in an interesting way.

So write short direction, look at the result, and respond to it. “Slower.” “She would not apologize here.” “Same beat, but he is already at the door.” “Too many adjectives in the second sentence.” This is editing, which you can do without training, rather than prompt engineering, which is a discipline you would have to learn before it paid off. A working writer already knows how to say what is wrong with a paragraph. Use that.

The other advantage of reacting is that your objections are specific and therefore reusable. “I don't like it” is not information. “Mara doesn't explain herself, ever” is a rule about a character that holds for the rest of the book, and the moment you say it out loud it becomes something a tool can carry forward — if it lands somewhere the tool will actually read. That last clause is the part beginners misjudge. OpenAI's help documentation separates two mechanisms in ChatGPT: saved memories you explicitly asked it to keep, and details inferred from past chats, which OpenAI says can change over time as ChatGPT updates what is more helpful to remember. OpenAI's own advice is to use saved memories for anything that must always be remembered. StoryProp runs that line down the middle of the product: what you agree to gets written down as a record the agent reads before it drafts, and what you don't stays a suggestion.

Guard your voice from the first exchange

A common regret after a first AI writing session is a manuscript that reads like nobody. This does not happen because a machine overwrote a strong voice. It happens because the writer had not yet decided what the voice was, and the default register — smooth, mid-Atlantic, slightly overwritten, fond of tricolons and em-dashes — filled the vacuum. Once fifteen pages exist in that register, everything new gets written to match it, including your own sentences.

The defense is early and unglamorous. In your first session, write a hundred words yourself. Badly is fine. Then hand them over and say: this is what the book sounds like, match the sentence length and the vocabulary level, do not add metaphors I would not use. You have just established a comparison point that every later draft can be measured against, and you can measure it, because you wrote the standard.

Be specific about mechanics rather than mood. “Literary” means nothing operationally. “Short declarative sentences, no semicolons, dialogue without attributive adverbs, characters do not describe their own emotions” means something an agent can actually apply, and means something you can check against the page in ten seconds. Style rules you can verify are the ones worth confirming. It also helps to say what you do not want, once, in plain terms. No weather at the top of scenes. No one gazes at anything. Nobody's breath catches. These sound petty written down, and they save you from the specific sludge that accumulates when a draft is generated to a generic notion of good prose.

The mush trap

There is a failure mode particular to the first week, and it is worth naming because it does not look like failure while it is happening. You get a paragraph. It is fine. You accept it and ask for the next one. That one is fine too. Twenty exchanges later you have three thousand words that are individually acceptable and collectively inert — no pressure, no particular voice, nothing at stake in any line. This is the mush trap, and the mechanism is that “fine” is a lower bar than “mine,” and it takes almost no energy to keep clearing it.

The trap is set by acceptance being cheaper than revision. Saying yes takes a keystroke. Saying “no, she is angrier than that, and she is not going to say so” takes thirty seconds of thought. Multiply across an afternoon and you can see how a draft ends up smooth and dead. Two habits break it. First, revise at least one thing in every exchange, even when the paragraph is acceptable — pick the weakest sentence and say why. This keeps you in the driver's seat as a matter of muscle memory rather than intention. Second, stop every few pages and read the whole run aloud. Mush is inaudible on screen and obvious in the mouth. You will hear the flatness in the second paragraph.

One more safeguard: work somewhere that keeps every version, so pushing back costs you nothing. StoryProp keeps every draft — each session, each pass — so you can compare two versions side by side, restore an earlier one without losing the later one, and undo a whole pass when an experiment goes nowhere. Fear of losing a decent paragraph is what makes people accept decent paragraphs. Take the fear away and you experiment more, and experimenting more is the whole cure.

What week one actually looks like

A realistic first week is not a chapter a day. It is a small number of short sessions in which you build a shared vocabulary with the tool and a clearer sense of your own book. Day one: one scene, small units, lots of pushback, a rough style statement written down. Day two: reread day one cold, cut what is generic, and notice which of your corrections you had to repeat — those are rules, so record them as rules.

By midweek you should be doing something more interesting than generating: asking questions. What does Mara want in this scene that she would not admit? Why does the chipped cup keep appearing? An agent with your records in front of it answers those in terms of your book rather than in general, and thinking out loud about your own story is often the highest-value thing you do all week.

By the end of the week the useful output is not word count. The Science Fiction and Fantasy Writers Association's Nebula rules put the floor for a novel at 40,000 words, and you will be nowhere near it, which is exactly right for seven days. What you should have is a page or two you would actually show someone, a short list of confirmed style rules, and a handful of character facts you now consider settled. That is a foundation. If you want to see how the pieces — records, sessions, revision — fit together, the features overview covers the mechanics, and when you are ready to scale this method up to a full manuscript, writing a novel with AI picks up where the first week ends.

The shape of the thing is this: what you are building in the first week is the habit of saying what you want in language precise enough to be acted on — which is, inconveniently and usefully, the same skill that makes the writing good in the first place. Nobody starts out with it. Everyone who keeps going ends up with it, and a week of eighty-word arguments is a fast way to get there.

The first session goes better when what you decided in it is still true in the tenth. That is what StoryProp is built for: say a rule once and it becomes a record the agent reads before it writes, so session two starts on the book instead of on the setup.

← All posts